diff --git a/evaluation/ner/direct_llm_call/neuroscience_ner_openai.py b/evaluation/ner/direct_llm_call/neuroscience_ner_openai.py
new file mode 100644
index 0000000..62e70ad
--- /dev/null
+++ b/evaluation/ner/direct_llm_call/neuroscience_ner_openai.py
@@ -0,0 +1,229 @@
+#!/usr/bin/env python3
+"""Run neuroscience-wide NER on an uploaded file via an OpenAI model.
+
+Usage:
+    python neuroscience_ner_openai.py --file paper.pdf --model gpt-5.5
+"""
+import argparse
+import json
+import os
+import re
+import sys
+from datetime import datetime
+
+from dotenv import load_dotenv
+from openai import OpenAI
+
+load_dotenv()
+
+# Default system/extractor prompt shipped alongside this script. Override with
+# --prompt-file to test alternative prompts.
+DEFAULT_PROMPT_FILE = os.path.join(
+    os.path.dirname(os.path.abspath(__file__)),
+    "prompts",
+    "extractor_neuroscience_ner.txt",
+)
+
+USER_PROMPT = """\
+INPUT TEXT:
+The text to process is the attached file. Treat the full extracted text of
+the attached document as the INPUT TEXT.
+
+METADATA (paper_title / doi / source_path) — populate `source_metadata` from
+this; do NOT repeat on every entity:
+{metadata_json}
+"""
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Run neuroscience-wide NER on an uploaded file via an OpenAI model."
+    )
+    parser.add_argument("--file", "-f", required=True, help="Path to the file to upload.")
+    parser.add_argument("--model", "-m", required=True, help="Model name, e.g. gpt-5.5.")
+    parser.add_argument(
+        "--prompt-file",
+        "-p",
+        default=DEFAULT_PROMPT_FILE,
+        help=f"Path to the system/extractor prompt file. Default: {DEFAULT_PROMPT_FILE}",
+    )
+    parser.add_argument(
+        "--metadata",
+        default="{}",
+        help='JSON string with paper_title / doi / source_path. Default: "{}".',
+    )
+    parser.add_argument(
+        "--output-dir",
+        "-o",
+        default=".",
+        help="Directory to write the output JSON file into. Default: current dir.",
+    )
+    parser.add_argument(
+        "--temperature",
+        "-t",
+        type=float,
+        default=None,
+        help="Sampling temperature for reproducible runs (e.g. 0). "
+        "Omitted by default; some reasoning models reject this.",
+    )
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=None,
+        help="Sampling seed for reproducible runs. Omitted by default; "
+        "some reasoning models reject this.",
+    )
+    parser.add_argument(
+        "--verbose",
+        "-v",
+        action="store_true",
+        help="Print step-by-step progress and streaming updates. Default: quiet.",
+    )
+    args = parser.parse_args()
+
+    def vprint(*a, **k):
+        """Print only when --verbose is set."""
+        if args.verbose:
+            print(*a, **k)
+
+    # Parse the metadata once. We both pass it to the model (for context) and
+    # stamp it authoritatively into the saved JSON below, so the output's
+    # source_metadata is correct regardless of what the model echoes.
+    try:
+        metadata = json.loads(args.metadata)
+        if not isinstance(metadata, dict):
+            raise ValueError("--metadata must be a JSON object")
+    except (json.JSONDecodeError, ValueError) as exc:
+        parser.error(f"invalid --metadata: {exc}")
+
+    # Default source_path to the input filename when not explicitly provided.
+    metadata.setdefault("source_path", args.file)
+
+    vprint(f"Loading system prompt from: {args.prompt_file}")
+    with open(args.prompt_file, "r", encoding="utf-8") as fh:
+        system_prompt = fh.read()
+
+    vprint("Initializing OpenAI client...")
+    # Generous read timeout for very long generations; fail rather than hang forever.
+    client = OpenAI(timeout=1800.0, max_retries=2)
+
+    vprint(f"Uploading file: {args.file} ...")
+    with open(args.file, "rb") as fh:
+        uploaded = client.files.create(file=fh, purpose="user_data")
+    vprint(f"  uploaded (file_id={uploaded.id})")
+
+    vprint(f"Sending request to model '{args.model}' (streaming; this may take a while)...")
+    request_input = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "input_file", "file_id": uploaded.id},
+                {
+                    "type": "input_text",
+                    "text": USER_PROMPT.format(metadata_json=json.dumps(metadata)),
+                },
+            ],
+        }
+    ]
+
+    # Only send sampling controls when explicitly requested; some reasoning
+    # models reject temperature/seed outright.
+    request_kwargs = {}
+    if args.temperature is not None:
+        request_kwargs["temperature"] = args.temperature
+    if args.seed is not None:
+        request_kwargs["seed"] = args.seed
+    if request_kwargs:
+        vprint(f"  sampling controls: {request_kwargs}")
+
+    chunks = []
+    chars = 0
+    next_report = 2000  # print a progress line every ~2000 chars
+    with client.responses.stream(
+        model=args.model,
+        instructions=system_prompt,
+        input=request_input,
+        **request_kwargs,
+    ) as stream:
+        for event in stream:
+            if event.type == "response.output_text.delta":
+                chunks.append(event.delta)
+                chars += len(event.delta)
+                if chars >= next_report:
+                    vprint(f"  ...streaming, {chars} chars received so far")
+                    next_report += 2000
+            elif event.type == "error":
+                print(f"  stream error: {event.error}")
+        final_response = stream.get_final_response()  # surfaces any terminal API error
+
+    output_text = "".join(chunks)
+    vprint(f"  response complete — {chars} chars total.")
+
+    # Detect truncation: a low entity count is often an output-token cutoff
+    # rather than the model deciding it was done.
+    status = getattr(final_response, "status", None)
+    if status == "incomplete":
+        reason = getattr(getattr(final_response, "incomplete_details", None), "reason", None)
+        print(
+            f"WARNING: response is INCOMPLETE (reason: {reason}). "
+            "Output was truncated — entity count is an undercount. "
+            "Consider raising the model's output-token limit or chunking the input."
+        )
+
+    # Build a filesystem-safe filename with timestamp and model.
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+    safe_model = re.sub(r"[^A-Za-z0-9._-]+", "-", args.model)
+    out_name = f"neuroscience_ner_{safe_model}_{timestamp}.json"
+    out_path = os.path.join(args.output_dir, out_name)
+
+    # Persist parsed JSON when possible; otherwise wrap the raw text.
+    vprint("Parsing model output as JSON...")
+    try:
+        payload = json.loads(output_text)
+        n_entities = len(payload.get("entities", [])) if isinstance(payload, dict) else 0
+        n_terms = len(payload.get("key_terms", [])) if isinstance(payload, dict) else 0
+        vprint(f"  parsed OK — {n_entities} entities, {n_terms} key_terms.")
+    except json.JSONDecodeError:
+        print("  output was not valid JSON — wrapping raw text under 'raw_output'.")
+        payload = {"raw_output": output_text}
+
+    # Stamp metadata authoritatively, overriding whatever the model echoed.
+    if isinstance(payload, dict):
+        payload["source_metadata"] = metadata
+        vprint(f"  stamped source_metadata: {metadata}")
+
+        # Compute extraction statistics and stamp them into the metadata.
+        entities = payload.get("entities", []) or []
+        label_counts = {}
+        for ent in entities:
+            if isinstance(ent, dict):
+                label = ent.get("label", "Unknown")
+                label_counts[label] = label_counts.get(label, 0) + 1
+        stats = {
+            "model": args.model,
+            "prompt_file": args.prompt_file,
+            "temperature": args.temperature,
+            "seed": args.seed,
+            "response_status": status,
+            "extracted_at": timestamp,
+            "total_entities": len(entities),
+            "entities_by_label": dict(
+                sorted(label_counts.items(), key=lambda kv: kv[1], reverse=True)
+            ),
+        }
+        payload["source_metadata"]["statistics"] = stats
+        print(
+            f"Extracted {stats['total_entities']} entities "
+            f"across {len(label_counts)} labels."
+        )
+
+    os.makedirs(args.output_dir, exist_ok=True)
+    vprint(f"Writing output to {out_path} ...")
+    with open(out_path, "w", encoding="utf-8") as fh:
+        json.dump(payload, fh, ensure_ascii=False, indent=2)
+
+    print(f"Done. Wrote output to {out_path}")
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/evaluation/ner/direct_llm_call/neuroscience_ner_openrouter.py b/evaluation/ner/direct_llm_call/neuroscience_ner_openrouter.py
new file mode 100644
index 0000000..1b1a8f4
--- /dev/null
+++ b/evaluation/ner/direct_llm_call/neuroscience_ner_openrouter.py
@@ -0,0 +1,464 @@
+#!/usr/bin/env python3
+"""Run neuroscience-wide NER on a local PDF via an OpenRouter model.
+
+Uses the OpenAI SDK pointed at OpenRouter, sending the PDF as a base64 data
+URL and configuring PDF parsing through the OpenRouter file-parser plugin.
+
+Usage:
+    python neuroscience_ner_openrouter.py --file paper.pdf --model openai/gpt-5.5
+"""
+import argparse
+import base64
+import json
+import os
+import re
+import sys
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+
+from dotenv import load_dotenv
+from openai import OpenAI
+
+load_dotenv()
+
+OPENROUTER_BASE_URL = "https://openrouter.ai/api/v1"
+
+
+class GrobidClient:
+    """
+    Thin wrapper around the Grobid HTTP API for full-text PDF parsing.
+
+    Grobid is expected to be running as a service (Docker image
+    ``lfoppiano/grobid`` on port 8070 by default). This client POSTs a PDF
+    to the ``processFulltextDocument`` endpoint and returns the resulting
+    TEI XML string.
+
+    Parameters
+    ----------
+    base_url:
+        Root URL of the Grobid service, e.g. ``http://localhost:8070``.
+    timeout:
+        HTTP timeout in seconds. Large papers (>30 pages) can take 60+ seconds
+        to parse so the default is set generously.
+    consolidate_citations:
+        Grobid flag — set to 0 (off), 1 (consolidate via CrossRef), or 2
+        (consolidate against local biblio-glutton service).
+    """
+
+    FULLTEXT_ENDPOINT = "/api/processFulltextDocument"
+    ISALIVE_ENDPOINT = "/api/isalive"
+
+    def __init__(
+        self,
+        base_url: str = "http://localhost:8070",
+        timeout: float = 300.0,
+        consolidate_citations: int = 0,
+    ) -> None:
+        try:
+            import requests as requests_mod
+        except ImportError as exc:  # pragma: no cover
+            raise ImportError(
+                "The `requests` package is required for GrobidClient. "
+                "Install with: uv add requests"
+            ) from exc
+
+        self._requests: Any = requests_mod
+        self.base_url = base_url.rstrip("/")
+        self.timeout = timeout
+        self.consolidate_citations = consolidate_citations
+
+    def is_alive(self) -> bool:
+        """Return True if the Grobid server responds at /api/isalive."""
+        req = self._requests
+        try:
+            r = req.get(self.base_url + self.ISALIVE_ENDPOINT, timeout=5)
+            return r.status_code == 200
+        except req.RequestException:
+            return False
+
+    def process_fulltext(self, pdf_path: Path) -> str:
+        """
+        POST *pdf_path* to Grobid's full-text endpoint and return TEI XML.
+
+        Parameters
+        ----------
+        pdf_path:
+            Path to the source PDF on disk.
+
+        Returns
+        -------
+        str
+            UTF-8 TEI XML string as returned by Grobid.
+
+        Raises
+        ------
+        FileNotFoundError
+            If *pdf_path* does not exist.
+        RuntimeError
+            If Grobid returns a non-2xx status, or the request fails.
+        """
+        pdf_path = Path(pdf_path)
+        if not pdf_path.is_file():
+            raise FileNotFoundError(f"PDF not found: {pdf_path}")
+
+        url = self.base_url + self.FULLTEXT_ENDPOINT
+        req = self._requests
+        try:
+            with pdf_path.open("rb") as fh:
+                resp = req.post(
+                    url,
+                    files={"input": (pdf_path.name, fh, "application/pdf")},
+                    data={"consolidateCitations": str(self.consolidate_citations)},
+                    timeout=self.timeout,
+                )
+        except req.RequestException as exc:
+            raise RuntimeError(
+                f"Grobid request to {url} failed: {exc}. "
+                f"Is the Grobid server running and reachable?"
+            ) from exc
+
+        if not resp.ok:
+            raise RuntimeError(
+                f"Grobid returned HTTP {resp.status_code} for {pdf_path.name}: "
+                f"{resp.text[:200]}"
+            )
+        return resp.text
+
+# Optional attribution headers for openrouter.ai rankings.
+EXTRA_HEADERS = {
+    "HTTP-Referer": "https://github.com/sensein/structsense",
+    "X-Title": "structsense-ner-eval",
+}
+
+# Default system/extractor prompt shipped alongside this script. Override with
+# --prompt-file to test alternative prompts. Shared with the OpenAI variant.
+DEFAULT_PROMPT_FILE = os.path.join(
+    os.path.dirname(os.path.abspath(__file__)),
+    "prompts",
+    "extractor_neuroscience_ner.txt",
+)
+
+# Used when the PDF is sent as an attachment and parsed by OpenRouter.
+USER_PROMPT_ATTACHED = """\
+INPUT TEXT:
+The text to process is the attached file. Treat the full extracted text of
+the attached document as the INPUT TEXT.
+
+METADATA (paper_title / doi / source_path) — populate `source_metadata` from
+this; do NOT repeat on every entity:
+{metadata_json}
+"""
+
+# Used when pre-parsed text (e.g. XML) is inlined directly into the prompt.
+USER_PROMPT_INLINE = """\
+INPUT TEXT:
+<<<
+{input_text}
+>>>
+
+METADATA (paper_title / doi / source_path) — populate `source_metadata` from
+this; do NOT repeat on every entity:
+{metadata_json}
+"""
+
+
+def encode_pdf_to_data_url(pdf_path):
+    """Read a local PDF and return it as a base64 data URL."""
+    with open(pdf_path, "rb") as fh:
+        encoded = base64.b64encode(fh.read()).decode("utf-8")
+    return f"data:application/pdf;base64,{encoded}"
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Run neuroscience-wide NER on a local PDF via an OpenRouter model."
+    )
+    parser.add_argument("--file", "-f", required=True, help="Path to the PDF to process.")
+    parser.add_argument(
+        "--model", "-m", required=True, help="OpenRouter model, e.g. openai/gpt-5.5."
+    )
+    parser.add_argument(
+        "--prompt-file",
+        "-p",
+        default=DEFAULT_PROMPT_FILE,
+        help=f"Path to the system/extractor prompt file. Default: {DEFAULT_PROMPT_FILE}",
+    )
+    parser.add_argument(
+        "--pdf-engine",
+        default="cloudflare-ai",
+        choices=["mistral-ocr", "cloudflare-ai", "native"],
+        help="OpenRouter file-parser PDF engine. Default: cloudflare-ai. "
+        "Ignored when --grobid is set.",
+    )
+    parser.add_argument(
+        "--grobid",
+        action="store_true",
+        help="Parse the PDF locally with Grobid to TEI XML before sending to "
+        "the LLM (inlined as text), instead of using OpenRouter's PDF parser.",
+    )
+    parser.add_argument(
+        "--grobid-url",
+        default="http://localhost:8070",
+        help="Base URL of the Grobid service. Default: http://localhost:8070.",
+    )
+    parser.add_argument(
+        "--metadata",
+        default="{}",
+        help='JSON string with paper_title / doi / source_path. Default: "{}".',
+    )
+    parser.add_argument(
+        "--output-dir",
+        "-o",
+        default=".",
+        help="Directory to write the output JSON file into. Default: current dir.",
+    )
+    parser.add_argument(
+        "--temperature",
+        "-t",
+        type=float,
+        default=None,
+        help="Sampling temperature for reproducible runs (e.g. 0). "
+        "Omitted by default; some reasoning models reject this.",
+    )
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=None,
+        help="Sampling seed for reproducible runs. Omitted by default; "
+        "some reasoning models reject this.",
+    )
+    parser.add_argument(
+        "--verbose",
+        "-v",
+        action="store_true",
+        help="Print step-by-step progress and streaming updates. Default: quiet.",
+    )
+    args = parser.parse_args()
+
+    def vprint(*a, **k):
+        """Print only when --verbose is set."""
+        if args.verbose:
+            print(*a, **k)
+
+    # Parse the metadata once. We both pass it to the model (for context) and
+    # stamp it authoritatively into the saved JSON below, so the output's
+    # source_metadata is correct regardless of what the model echoes.
+    try:
+        metadata = json.loads(args.metadata)
+        if not isinstance(metadata, dict):
+            raise ValueError("--metadata must be a JSON object")
+    except (json.JSONDecodeError, ValueError) as exc:
+        parser.error(f"invalid --metadata: {exc}")
+
+    # Default source_path to the input filename when not explicitly provided.
+    metadata.setdefault("source_path", args.file)
+
+    api_key = os.getenv("OPENROUTER_API_KEY")
+    if not api_key:
+        parser.error("OPENROUTER_API_KEY is not set (check your .env).")
+
+    vprint(f"Loading system prompt from: {args.prompt_file}")
+    with open(args.prompt_file, "r", encoding="utf-8") as fh:
+        system_prompt = fh.read()
+
+    vprint("Initializing OpenRouter client...")
+    # Generous read timeout for very long generations; fail rather than hang forever.
+    client = OpenAI(
+        base_url=OPENROUTER_BASE_URL,
+        api_key=api_key,
+        timeout=1800.0,
+        max_retries=2,
+    )
+
+    # Three input modes:
+    #   - PDF + --grobid : parse locally to TEI XML, then inline as text.
+    #   - PDF (default)   : send as attachment, parsed by OpenRouter's plugin.
+    #   - non-PDF         : pre-parsed text (XML, plain text) inlined directly.
+    is_pdf = args.file.lower().endswith(".pdf")
+    metadata_json = json.dumps(metadata)
+
+    if is_pdf and args.grobid:
+        vprint(f"Parsing PDF with Grobid at {args.grobid_url} ...")
+        grobid = GrobidClient(base_url=args.grobid_url)
+        if not grobid.is_alive():
+            parser.error(
+                f"Grobid service is not reachable at {args.grobid_url}. "
+                "Start it (e.g. `docker run -p 8070:8070 lfoppiano/grobid`) "
+                "or pass --grobid-url."
+            )
+        input_text = grobid.process_fulltext(Path(args.file))
+        vprint(f"  Grobid returned {len(input_text)} chars of TEI XML.")
+        input_mode = "grobid_tei"
+        messages = [
+            {"role": "system", "content": system_prompt},
+            {
+                "role": "user",
+                "content": USER_PROMPT_INLINE.format(
+                    input_text=input_text, metadata_json=metadata_json
+                ),
+            },
+        ]
+        plugins = None
+    elif is_pdf:
+        vprint(f"Encoding PDF: {args.file} ...")
+        data_url = encode_pdf_to_data_url(args.file)
+        input_mode = "pdf_attachment"
+        messages = [
+            {"role": "system", "content": system_prompt},
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "text",
+                        "text": USER_PROMPT_ATTACHED.format(metadata_json=metadata_json),
+                    },
+                    {
+                        "type": "file",
+                        "file": {
+                            "filename": os.path.basename(args.file),
+                            "file_data": data_url,
+                        },
+                    },
+                ],
+            },
+        ]
+        plugins = [{"id": "file-parser", "pdf": {"engine": args.pdf_engine}}]
+    else:
+        vprint(f"Reading pre-parsed text file: {args.file} ...")
+        with open(args.file, "r", encoding="utf-8") as fh:
+            input_text = fh.read()
+        vprint(f"  read {len(input_text)} chars.")
+        input_mode = "inline_text"
+        messages = [
+            {"role": "system", "content": system_prompt},
+            {
+                "role": "user",
+                "content": USER_PROMPT_INLINE.format(
+                    input_text=input_text, metadata_json=metadata_json
+                ),
+            },
+        ]
+        plugins = None
+
+    # Only send sampling controls when explicitly requested; some reasoning
+    # models reject temperature/seed outright.
+    request_kwargs = {}
+    if args.temperature is not None:
+        request_kwargs["temperature"] = args.temperature
+    if args.seed is not None:
+        request_kwargs["seed"] = args.seed
+    if request_kwargs:
+        vprint(f"  sampling controls: {request_kwargs}")
+
+    # The file-parser plugin only applies to PDF attachments.
+    if plugins is not None:
+        request_kwargs["extra_body"] = {"plugins": plugins}
+
+    if input_mode == "pdf_attachment":
+        source_desc = f"pdf-engine={args.pdf_engine}"
+    elif input_mode == "grobid_tei":
+        source_desc = "grobid TEI"
+    else:
+        source_desc = "inline text"
+    vprint(
+        f"Sending request to model '{args.model}' "
+        f"({source_desc}, streaming; this may take a while)..."
+    )
+    chunks = []
+    chars = 0
+    next_report = 2000  # print a progress line every ~2000 chars
+    finish_reason = None
+    stream = client.chat.completions.create(
+        model=args.model,
+        messages=messages,
+        extra_headers=EXTRA_HEADERS,
+        stream=True,
+        **request_kwargs,
+    )
+    for chunk in stream:
+        if not chunk.choices:
+            continue
+        choice = chunk.choices[0]
+        delta = getattr(choice.delta, "content", None)
+        if delta:
+            chunks.append(delta)
+            chars += len(delta)
+            if chars >= next_report:
+                vprint(f"  ...streaming, {chars} chars received so far")
+                next_report += 2000
+        if choice.finish_reason:
+            finish_reason = choice.finish_reason
+
+    output_text = "".join(chunks)
+    vprint(f"  response complete — {chars} chars total (finish_reason={finish_reason}).")
+
+    # Detect truncation: a low entity count is often an output-token cutoff
+    # rather than the model deciding it was done.
+    if finish_reason == "length":
+        print(
+            "WARNING: response was truncated (finish_reason='length'). "
+            "Entity count is an undercount. "
+            "Consider raising the model's output-token limit or chunking the input."
+        )
+
+    # Build a filesystem-safe filename with timestamp and model.
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+    safe_model = re.sub(r"[^A-Za-z0-9._-]+", "-", args.model)
+    out_name = f"neuroscience_ner_{safe_model}_{timestamp}.json"
+    out_path = os.path.join(args.output_dir, out_name)
+
+    # Persist parsed JSON when possible; otherwise wrap the raw text.
+    vprint("Parsing model output as JSON...")
+    try:
+        payload = json.loads(output_text)
+        n_entities = len(payload.get("entities", [])) if isinstance(payload, dict) else 0
+        n_terms = len(payload.get("key_terms", [])) if isinstance(payload, dict) else 0
+        vprint(f"  parsed OK — {n_entities} entities, {n_terms} key_terms.")
+    except json.JSONDecodeError:
+        print("  output was not valid JSON — wrapping raw text under 'raw_output'.")
+        payload = {"raw_output": output_text}
+
+    # Stamp metadata authoritatively, overriding whatever the model echoed.
+    if isinstance(payload, dict):
+        payload["source_metadata"] = metadata
+        vprint(f"  stamped source_metadata: {metadata}")
+
+        # Compute extraction statistics and stamp them into the metadata.
+        entities = payload.get("entities", []) or []
+        label_counts = {}
+        for ent in entities:
+            if isinstance(ent, dict):
+                label = ent.get("label", "Unknown")
+                label_counts[label] = label_counts.get(label, 0) + 1
+        stats = {
+            "model": args.model,
+            "prompt_file": args.prompt_file,
+            "input_mode": input_mode,
+            "pdf_engine": args.pdf_engine if input_mode == "pdf_attachment" else None,
+            "grobid_url": args.grobid_url if input_mode == "grobid_tei" else None,
+            "temperature": args.temperature,
+            "seed": args.seed,
+            "finish_reason": finish_reason,
+            "extracted_at": timestamp,
+            "total_entities": len(entities),
+            "entities_by_label": dict(
+                sorted(label_counts.items(), key=lambda kv: kv[1], reverse=True)
+            ),
+        }
+        payload["source_metadata"]["statistics"] = stats
+        print(
+            f"Extracted {stats['total_entities']} entities "
+            f"across {len(label_counts)} labels."
+        )
+
+    os.makedirs(args.output_dir, exist_ok=True)
+    vprint(f"Writing output to {out_path} ...")
+    with open(out_path, "w", encoding="utf-8") as fh:
+        json.dump(payload, fh, ensure_ascii=False, indent=2)
+
+    print(f"Done. Wrote output to {out_path}")
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_gpt-5.5_20260628_231635.json b/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_gpt-5.5_20260628_231635.json
new file mode 100644
index 0000000..3c0eba2
--- /dev/null
+++ b/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_gpt-5.5_20260628_231635.json
@@ -0,0 +1,2745 @@
+{
+  "source_metadata": {
+    "paper_title": null,
+    "doi": null,
+    "source_path": null
+  },
+  "entities": [
+    {
+      "entity": "Higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "Dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "dimensionality reduction",
+      "label": "Method",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "recurrent connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "recurrent neural networks",
+      "label": "ComputationalModel",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "sensory responses",
+      "label": "Measurement",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "sensory responses",
+      "label": "Measurement",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "Cognitive functions",
+      "label": "Phenomenon",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "excitatory and inhibitory neurons",
+      "label": "CellType",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "cortical circuits",
+      "label": "NeuralCircuit",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "neural populations",
+      "label": "CellType",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "excitation",
+      "label": "Phenomenon",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "inhibition",
+      "label": "Phenomenon",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "dynamical system",
+      "label": "ComputationalModel",
+      "sentence": "Because these circuit models usually assume a relatively simple connectivity structure, their connectivity can be directly related to a dynamical system description of computations supporting cognitive task execution5,6.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "neural trajectories",
+      "label": "Measurement",
+      "sentence": "Thus, these models explicitly specify a circuit mechanism in the connectivity structure that gives rise to a dynamical mechanism controlling the flow of neural trajectories to implement task computations.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "excitation–inhibition balance",
+      "label": "Phenomenon",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation–inhibition balance12) and thus can be causally validated in experiments13–15.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of single-neuron responses in the cortex.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "cortex",
+      "label": "BrainRegion",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of single-neuron responses in the cortex.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "Single neurons",
+      "label": "CellType",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "heterogeneous tuning",
+      "label": "Measurement",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "recurrent neural network",
+      "label": "ComputationalModel",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "neural activity data",
+      "label": "Measurement",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks30.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks30.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "fixed points",
+      "label": "DynamicalSystemFeature",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "linearized dynamics",
+      "label": "DynamicalSystemFeature",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "dynamical system",
+      "label": "ComputationalModel",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Page 1"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Thus, high-dimensional RNNs currently serve as intermediate mechanistic models of heterogeneous neural responses, which yield dynamical mechanisms but leave the underlying circuit mechanisms unknown.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "Thus, high-dimensional RNNs currently serve as intermediate mechanistic models of heterogeneous neural responses, which yield dynamical mechanisms but leave the underlying circuit mechanisms unknown.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas30,36.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "joint population activity",
+      "label": "Measurement",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas30,36.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "brain areas",
+      "label": "BrainRegion",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas30,36.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "neural population activity",
+      "label": "Measurement",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "latent variables",
+      "label": "ComputationalVariable",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "unsupervised methods",
+      "label": "Method",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution26,37–40.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "latent variables",
+      "label": "ComputationalVariable",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution26,37–40.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "targeted dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "neural representations",
+      "label": "Measurement",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "neural population activity",
+      "label": "Measurement",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "However, unlike RNNs and circuit models, these correlation-based methods do not incorporate recurrent interactions among task variables, which implement computations necessary to solve the task.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "correlation-based methods",
+      "label": "Method",
+      "sentence": "However, unlike RNNs and circuit models, these correlation-based methods do not incorporate recurrent interactions among task variables, which implement computations necessary to solve the task.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC’s role in context-dependent decision-making.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC’s role in context-dependent decision-making.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "Context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus–response mappings.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "inhibition",
+      "label": "Phenomenon",
+      "sentence": "Most circuit models hypothesize a relatively simple mechanism based on inhibition of sensory representations irrelevant in the current context8–10,44–46.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "Measurement",
+      "sentence": "Most circuit models hypothesize a relatively simple mechanism based on inhibition of sensory representations irrelevant in the current context8–10,44–46.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "sensory responses",
+      "label": "Measurement",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "These results suggest that RNNs and conceivably the PFC implement qualitatively novel task solutions that do not exist in small circuits and emerge only in high-dimensional networks19,47,48.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "These results suggest that RNNs and conceivably the PFC implement qualitatively novel task solutions that do not exist in small circuits and emerge only in high-dimensional networks19,47,48.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "dimensionality reduction",
+      "label": "Method",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "latent circuit connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "The model simultaneously infers a low-dimensional latent circuit connectivity generating task-relevant dynamics and heterogeneous mixing of these dynamics in single-neuron responses.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "The model simultaneously infers a low-dimensional latent circuit connectivity generating task-relevant dynamics and heterogeneous mixing of these dynamics in single-neuron responses.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "inhibition",
+      "label": "Phenomenon",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "Measurement",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "We validated this mechanism by confirming the behavioral effects of patterned perturbations of the RNN activity and connectivity predicted by the latent circuit model.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We validated this mechanism by confirming the behavioral effects of patterned perturbations of the RNN activity and connectivity predicted by the latent circuit model.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "neural recordings",
+      "label": "Method",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "RNN perturbations",
+      "label": "Method",
+      "sentence": "Using RNN perturbations, we show that dimensionality reduction methods that do not incorporate causal interactions among latent variables are biased toward uncovering behaviorally irrelevant representations.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Using RNN perturbations, we show that dimensionality reduction methods that do not incorporate causal interactions among latent variables are biased toward uncovering behaviorally irrelevant representations.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "latent variables",
+      "label": "ComputationalVariable",
+      "sentence": "Using RNN perturbations, we show that dimensionality reduction methods that do not incorporate causal interactions among latent variables are biased toward uncovering behaviorally irrelevant representations.",
+      "paper_location": "Page 2"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "Our results show that high-dimensional networks use low-dimensional circuit mechanisms, establish the feasibility of inferring these mechanisms from neural response data and open new possibilities of causally validating circuit mechanisms in perturbation experiments.",
+      "paper_location": "Page 3"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "single-neuron heterogeneity",
+      "label": "Measurement",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as y = Qx, (1)",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "neurons",
+      "label": "CellType",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as y = Qx, (1)",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "latent variables",
+      "label": "ComputationalVariable",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as y = Qx, (1)",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "ReLU",
+      "label": "ActivationFunction",
+      "sentence": "The latent variables x are constrained to be nodes in a neural circuit with dynamics ẋ= −x + f(wrecx + winu), (2) where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "rectified linear",
+      "label": "ActivationFunction",
+      "sentence": "The latent variables x are constrained to be nodes in a neural circuit with dynamics ẋ= −x + f(wrecx + winu), (2) where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "latent nodes",
+      "label": "ComputationalVariable",
+      "sentence": "The latent nodes interact via the recurrent connectivity wrec and receive external task inputs u through the input connectivity win.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "recurrent connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "The latent nodes interact via the recurrent connectivity wrec and receive external task inputs u through the input connectivity win.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity wout: z = woutx.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "circuit activity",
+      "label": "Measurement",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity wout: z = woutx.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "We infer the latent circuit parameters (Q, wrec, win and wout) from neural activity y by minimizing the loss function L = ∑k,t∥y − Qx∥2 + ∥z − woutx∥2, where k and t index trials and time within a trial, respectively (Methods).",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "latent nodes",
+      "label": "ComputationalVariable",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "Thus, the latent circuit provides a dimensionality reduction that incorporates an explicit mechanistic hypothesis for how the resulting low-dimensional dynamics are generated.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "dimensionality reduction",
+      "label": "Method",
+      "sentence": "Thus, the latent circuit provides a dimensionality reduction that incorporates an explicit mechanistic hypothesis for how the resulting low-dimensional dynamics are generated.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "low-dimensional dynamics",
+      "label": "DynamicalSystemFeature",
+      "sentence": "Thus, the latent circuit provides a dimensionality reduction that incorporates an explicit mechanistic hypothesis for how the resulting low-dimensional dynamics are generated.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "mixed selectivity",
+      "label": "Phenomenon",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "cortex",
+      "label": "BrainRegion",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We consider RNNs with dynamics ẏ= −y + f(Wrec y + Winu), (4) where Wrec and Win are the recurrent and input connectivity matrices, respectively (Methods).",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "Using the fact that the vector fields of the RNN and latent circuit are piecewise-linear dynamical systems, we derive a relationship between their recurrent and input connectivity matrices (Methods): QTWrecQ = wrec, QTWin = win.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Using the fact that the vector fields of the RNN and latent circuit are piecewise-linear dynamical systems, we derive a relationship between their recurrent and input connectivity matrices (Methods): QTWrecQ = wrec, QTWin = win.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "piecewise-linear dynamical systems",
+      "label": "ComputationalModel",
+      "sentence": "Using the fact that the vector fields of the RNN and latent circuit are piecewise-linear dynamical systems, we derive a relationship between their recurrent and input connectivity matrices (Methods): QTWrecQ = wrec, QTWin = win.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "latent circuit connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "The relation Eq. (5) shows that the latent circuit connectivity wrec is a low-rank structure in the connectivity of a high-dimensional network, which captures interactions among the latent variables defined by the columns of Q.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "latent variables",
+      "label": "ComputationalVariable",
+      "sentence": "The relation Eq. (5) shows that the latent circuit connectivity wrec is a low-rank structure in the connectivity of a high-dimensional network, which captures interactions among the latent variables defined by the columns of Q.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "The relation between connectivity matrices has the powerful consequence that we can validate the latent circuit mechanism directly in the RNN connectivity.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "First, if the latent circuit faithfully describes the mechanism operating in the RNN, by conjugating the RNN connectivity matrix with Q (Eq. (5)), we expect to find low-dimensional connectivity structure similar to the latent circuit connectivity.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "First, if the latent circuit faithfully describes the mechanism operating in the RNN, by conjugating the RNN connectivity matrix with Q (Eq. (5)), we expect to find low-dimensional connectivity structure similar to the latent circuit connectivity.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "First, if the latent circuit faithfully describes the mechanism operating in the RNN, by conjugating the RNN connectivity matrix with Q (Eq. (5)), we expect to find low-dimensional connectivity structure similar to the latent circuit connectivity.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "sensory stimulus",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "context cue",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "The RNN successfully learns the task; it makes choices according to the relevant stimulus and ignores the irrelevant stimulus in each context (Fig. 2c).",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalUnit",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "mixed selectivity",
+      "label": "Phenomenon",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "The structure in the RNN connectivity responsible for generating the correct behavioral outputs is not immediately obvious (Fig. 2d).",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We fitted the latent circuit model to the responses of RNN units and its output behavior during the task.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalUnit",
+      "sentence": "We fitted the latent circuit model to the responses of RNN units and its output behavior during the task.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalVariable",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "sensory color nodes",
+      "label": "ComputationalVariable",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "sensory motion nodes",
+      "label": "ComputationalVariable",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "choice nodes",
+      "label": "ComputationalVariable",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "This choice of the latent circuit dimensionality agrees with the dimensionality of RNN responses after training, which is usually close to the total number of inputs and outputs (the first eight principal components accounted for 97.9% of the total variance in RNN responses; Supplementary Fig. 2).",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "This choice of the latent circuit dimensionality agrees with the dimensionality of RNN responses after training, which is usually close to the total number of inputs and outputs (the first eight principal components accounted for 97.9% of the total variance in RNN responses; Supplementary Fig. 2).",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "coefficient of determination",
+      "label": "Measurement",
+      "sentence": "The fitted latent circuit model captured an overwhelming amount of variance in the RNN activity (coefficient of determination r2 = 0.96 on test data) and accurately matched projected RNN trajectories (Supplementary Fig. 3).",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "The fitted latent circuit model captured an overwhelming amount of variance in the RNN activity (coefficient of determination r2 = 0.96 on test data) and accurately matched projected RNN trajectories (Supplementary Fig. 3).",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "RNN trajectories",
+      "label": "Measurement",
+      "sentence": "The fitted latent circuit model captured an overwhelming amount of variance in the RNN activity (coefficient of determination r2 = 0.96 on test data) and accurately matched projected RNN trajectories (Supplementary Fig. 3).",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalVariable",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "excitatory connections",
+      "label": "NeuralConnectivity",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalVariable",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "inhibitory connections",
+      "label": "NeuralConnectivity",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalVariable",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "This pattern of connections from the context nodes to the sensory nodes implements a suppression mechanism that inhibits the irrelevant stimulus–response mapping in each context.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "irrelevant sensory representation",
+      "label": "Measurement",
+      "sentence": "Because the irrelevant sensory representation is suppressed, it does not drive the decision output.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "This suppression mechanism based on inhibition of irrelevant representations is qualitatively similar to mechanisms for context-dependent decision-making hypothesized in previous hand-crafted neural circuit models8,9.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "This suppression mechanism based on inhibition of irrelevant representations is qualitatively similar to mechanisms for context-dependent decision-making hypothesized in previous hand-crafted neural circuit models8,9.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "We verified that the suppression mechanism revealed in the latent circuit connectivity depended significantly on the RNN responses beyond the constraints imposed by the task alone (Extended Data Fig. 1 and Methods), suggesting that this mechanism reflects the dynamics of the RNN.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "We therefore proceeded to validate the inferred circuit mechanism directly in the RNN in three ways: in the RNN activity, in the RNN connectivity and by confirming behavioral effects of the RNN perturbations predicted by the latent circuit model.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We therefore proceeded to validate the inferred circuit mechanism directly in the RNN in three ways: in the RNN activity, in the RNN connectivity and by confirming behavioral effects of the RNN perturbations predicted by the latent circuit model.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "We therefore proceeded to validate the inferred circuit mechanism directly in the RNN in three ways: in the RNN activity, in the RNN connectivity and by confirming behavioral effects of the RNN perturbations predicted by the latent circuit model.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "We projected RNN responses onto the columns of Q, which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "RNN activity patterns",
+      "label": "Measurement",
+      "sentence": "We projected RNN responses onto the columns of Q, which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "We projected RNN responses onto the columns of Q, which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "entity": "choice axis",
+      "label": "ComputationalVariable",
+      "sentence": "Next, the choice axis is the difference of two columns of Q corresponding to the left and right choice nodes in the latent circuit.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "Projecting RNN activity onto the choice axis reveals trajectories separating according to choice regardless of context.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "motion axis",
+      "label": "ComputationalVariable",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "color axis",
+      "label": "ComputationalVariable",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "Projections of RNN activity onto the motion and color axes reveal representations of relevant sensory stimuli, whereas representations of irrelevant stimuli are suppressed.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "sensory stimuli",
+      "label": "Stimulus",
+      "sentence": "Projections of RNN activity onto the motion and color axes reveal representations of relevant sensory stimuli, whereas representations of irrelevant stimuli are suppressed.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "The suppression of irrelevant sensory representations in RNN activity is consistent with the inhibitory mechanism revealed in the latent circuit connectivity wrec.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "latent circuit connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "The suppression of irrelevant sensory representations in RNN activity is consistent with the inhibitory mechanism revealed in the latent circuit connectivity wrec.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "Second, we used the connectivity relationships Eq. (5) to directly validate the latent circuit mechanism in the RNN connectivity.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "latent circuit mechanism",
+      "label": "ComputationalModel",
+      "sentence": "Second, we used the connectivity relationships Eq. (5) to directly validate the latent circuit mechanism in the RNN connectivity.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "psychometric function",
+      "label": "Measurement",
+      "sentence": "Indeed, weakening these connections in the latent circuit produced the predicted behavioral effect in the psychometric function, visible as a rotation of the decision boundary on motion context trials (Fig. 4a).",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "Perturbations of the RNN connectivity along the corresponding pattern produced similar behavioral effects (Fig. 4b), confirming that this connectivity pattern implements suppression of irrelevant sensory representations in the RNN.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "Measurement",
+      "sentence": "Perturbations of the RNN connectivity along the corresponding pattern produced similar behavioral effects (Fig. 4b), confirming that this connectivity pattern implements suppression of irrelevant sensory representations in the RNN.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Perturbations of the RNN connectivity along the corresponding pattern produced similar behavioral effects (Fig. 4b), confirming that this connectivity pattern implements suppression of irrelevant sensory representations in the RNN.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "entity": "stimulus–response mappings",
+      "label": "TaskParadigm",
+      "sentence": "The second perturbation ‘turns off’ one of the stimulus–response mappings by weakening the excitatory connection from a sensory node to a choice node.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "sensory node",
+      "label": "ComputationalVariable",
+      "sentence": "The second perturbation ‘turns off’ one of the stimulus–response mappings by weakening the excitatory connection from a sensory node to a choice node.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "choice node",
+      "label": "ComputationalVariable",
+      "sentence": "The second perturbation ‘turns off’ one of the stimulus–response mappings by weakening the excitatory connection from a sensory node to a choice node.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Together, these results confirm that the RNN uses the suppression mechanism in which context representations inhibit irrelevant sensory representations.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "Together, these results confirm that the RNN uses the suppression mechanism in which context representations inhibit irrelevant sensory representations.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "Measurement",
+      "sentence": "Together, these results confirm that the RNN uses the suppression mechanism in which context representations inhibit irrelevant sensory representations.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "This mechanism is reflected in the low-dimensional dynamics revealed by projecting RNN activity onto axes defined by the latent circuit embedding Q.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "We identified this mechanism as a latent low-dimensional structure in the RNN connectivity and ultimately validated it by confirming behavioral effects of the RNN connectivity perturbations.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We next asked whether RNNs trained on the same task arrive at different circuit solutions for context-dependent decision-making.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "We next asked whether RNNs trained on the same task arrive at different circuit solutions for context-dependent decision-making.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "The latent circuit inference enables us to determine whether different RNNs use the same circuit mechanism.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit inference enables us to determine whether different RNNs use the same circuit mechanism.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "latent connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "Therefore, we assessed the similarity of task solutions in RNNs by comparing their low-dimensional latent connectivity.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "principal component analysis",
+      "label": "Method",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices wrec and projected the data onto the first two principal components, which accounted for 42% of total variance (Fig. 5a).",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices wrec and projected the data onto the first two principal components, which accounted for 42% of total variance (Fig. 5a).",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "Gaussian mixture model",
+      "label": "Method",
+      "sentence": "The projection reveals three major solution clusters, which we partitioned by fitting data with a Gaussian mixture model with three components.",
+      "paper_location": "Figure 5; Page 7"
+    },
+    {
+      "entity": "Mann–Whitney U-test",
+      "label": "StatisticalTest",
+      "sentence": "These latent models produced a significantly worse fit when the latent circuit and the target RNN were sampled from different clusters than from the same cluster (Extended Data Fig. 2m; one-sided Mann–Whitney U-test, U = 52, 593, P < 10−10), confirming differences in circuit mechanisms across clusters.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "entity": "inhibitory suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses19.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "dynamical selection vector mechanism",
+      "label": "ComputationalMechanism",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses19.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses19.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "sensory responses",
+      "label": "Measurement",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses19.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "Softplus",
+      "label": "ActivationFunction",
+      "sentence": "Furthermore, we found the same inhibitory suppression mechanism in RNNs trained without constraining their inputs to be orthogonal (Supplementary Fig. 6) and in RNNs with different biologically plausible nonlinearities (Extended Data Figs. 5 and 6).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "sensory responses",
+      "label": "Measurement",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "This conclusion was derived using dimensionality reduction methods that fit neural responses with regression models19,21,41–43 to find low-dimensional projections that best correlate with task variables (Fig. 1b).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "This conclusion was derived using dimensionality reduction methods that fit neural responses with regression models19,21,41–43 to find low-dimensional projections that best correlate with task variables (Fig. 1b).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "regression models",
+      "label": "Method",
+      "sentence": "This conclusion was derived using dimensionality reduction methods that fit neural responses with regression models19,21,41–43 to find low-dimensional projections that best correlate with task variables (Fig. 1b).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "We therefore sought to determine whether the latent circuit and regression models identify different task representations in the same PFC responses.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies19,21.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "PFC recordings",
+      "label": "Method",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies19,21.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies19,21.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "PFC neurons",
+      "label": "CellType",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "rhesus monkeys",
+      "label": "Species",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "condition-averaged PFC responses",
+      "label": "Measurement",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting 100 ms after the stimulus onset19 (Methods).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "stimulus onset",
+      "label": "StimulusTiming",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting 100 ms after the stimulus onset19 (Methods).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "The latent circuit model provided good fits of PFC responses projected onto the low-dimensional subspace spanned by the columns of the inferred embedding matrix Q for both monkeys (r2 = 0.88 and 0.76 on test data for monkeys A and F, respectively; Extended Data Figs. 7a and 8a).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "The latent circuit model provided good fits of PFC responses projected onto the low-dimensional subspace spanned by the columns of the inferred embedding matrix Q for both monkeys (r2 = 0.88 and 0.76 on test data for monkeys A and F, respectively; Extended Data Figs. 7a and 8a).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "The task subspace Q explained a smaller fraction of total variance in the PFC responses (11.0% and 7.0% for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC42 and reflects the high dimensionality of PFC activity19.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "The task subspace Q explained a smaller fraction of total variance in the PFC responses (11.0% and 7.0% for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC42 and reflects the high dimensionality of PFC activity19.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The task subspace Q explained a smaller fraction of total variance in the PFC responses (11.0% and 7.0% for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC42 and reflects the high dimensionality of PFC activity19.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "PFC activity",
+      "label": "Measurement",
+      "sentence": "The task subspace Q explained a smaller fraction of total variance in the PFC responses (11.0% and 7.0% for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC42 and reflects the high dimensionality of PFC activity19.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "stimulus representations",
+      "label": "Measurement",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "Mann–Whitney U-test",
+      "label": "StatisticalTest",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "monkey A",
+      "label": "Species",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "monkey F",
+      "label": "Species",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "entity": "PFC data",
+      "label": "Measurement",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "PFC activity",
+      "label": "Measurement",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "latent circuit connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "inhibitory connections",
+      "label": "NeuralConnectivity",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalVariable",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts19,21.",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "PFC activity",
+      "label": "Measurement",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts19,21.",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "PFC data",
+      "label": "Measurement",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts19,21.",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "regression methods",
+      "label": "Method",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts19,21.",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "error trials",
+      "label": "BehavioralEvent",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "motion axis",
+      "label": "ComputationalVariable",
+      "sentence": "Specifically, the suppression mechanism predicts that the representations of irrelevant stimuli (along the motion and color axes identified by the latent circuit model) should be less suppressed on error than correct trials.",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "color axes",
+      "label": "ComputationalVariable",
+      "sentence": "Specifically, the suppression mechanism predicts that the representations of irrelevant stimuli (along the motion and color axes identified by the latent circuit model) should be less suppressed on error than correct trials.",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "Method",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "Method",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "decoder axis",
+      "label": "ComputationalVariable",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "motion stimulus",
+      "label": "Stimulus",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Thus, as in the PFC, irrelevant sensory representations in our RNN appear not suppressed along the decoder axis, whereas they appear suppressed along the axis obtained from the latent circuit model (Fig. 7b).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Thus, as in the PFC, irrelevant sensory representations in our RNN appear not suppressed along the decoder axis, whereas they appear suppressed along the axis obtained from the latent circuit model (Fig. 7b).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "Measurement",
+      "sentence": "Thus, as in the PFC, irrelevant sensory representations in our RNN appear not suppressed along the decoder axis, whereas they appear suppressed along the axis obtained from the latent circuit model (Fig. 7b).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "How can we reconcile these qualitatively distinct perspectives on representations of irrelevant stimuli within the same RNN?",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "Method",
+      "sentence": "We hypothesized that the appearance of irrelevant stimulus representations is possible because the linear decoder compromises behavioral relevance for decoding accuracy.",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalUnit",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "activity patterns",
+      "label": "Measurement",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "Method",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "psychometric functions",
+      "label": "Measurement",
+      "sentence": "If the corresponding activity patterns are behaviorally relevant, we expect the stimulation to have a substantial effect on psychometric functions.",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "As expected, driving RNN activity along the motion axis of the latent circuit model shifted the decision boundary on motion context trials (Fig. 7d).",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "motion axis",
+      "label": "ComputationalVariable",
+      "sentence": "As expected, driving RNN activity along the motion axis of the latent circuit model shifted the decision boundary on motion context trials (Fig. 7d).",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "As expected, driving RNN activity along the motion axis of the latent circuit model shifted the decision boundary on motion context trials (Fig. 7d).",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "decoder axis",
+      "label": "ComputationalVariable",
+      "sentence": "By contrast, stimulation of the same magnitude along the decoder axis had little effect on the psychometric function (Fig. 7c).",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "psychometric function",
+      "label": "Measurement",
+      "sentence": "By contrast, stimulation of the same magnitude along the decoder axis had little effect on the psychometric function (Fig. 7c).",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "For context-dependent decision-making, we conclude that the dynamics revealed by the decoder have little behavioral relevance and thus do not invalidate the inhibitory mechanism identified by the latent circuit model.",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "For context-dependent decision-making, we conclude that the dynamics revealed by the decoder have little behavioral relevance and thus do not invalidate the inhibitory mechanism identified by the latent circuit model.",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "entity": "Single neurons",
+      "label": "CellType",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "cognitive functions",
+      "label": "Phenomenon",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "single-neuron heterogeneity",
+      "label": "Measurement",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "dimensionality reduction",
+      "label": "Method",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "latent variables",
+      "label": "ComputationalVariable",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "We show that low-dimensional circuit mechanisms can explain task-relevant dynamics in high-dimensional networks and establish feasibility of inferring these mechanisms from neural response data.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Our theory for interpreting the latent circuit connectivity as a low-rank connectivity in RNNs enables causally validating low-dimensional circuit mechanisms via activity and connectivity perturbations in high-dimensional networks.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "latent circuit connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "Our theory for interpreting the latent circuit connectivity as a low-rank connectivity in RNNs enables causally validating low-dimensional circuit mechanisms via activity and connectivity perturbations in high-dimensional networks.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "The latent circuit inference can be broadly applied to identify circuit mechanisms for different cognitive tasks from neural response data (Supplementary Fig. 7) and opens new possibilities for causally testing these mechanisms in future experiments.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "The latent circuit inference can be broadly applied to identify circuit mechanisms for different cognitive tasks from neural response data (Supplementary Fig. 7) and opens new possibilities for causally testing these mechanisms in future experiments.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "latent dynamical systems",
+      "label": "ComputationalModel",
+      "sentence": "Although previous studies extensively modeled neural responses using various latent dynamical systems, our work demonstrates feasibility of fitting neural responses with low-dimensional recurrent circuits.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "Although previous studies extensively modeled neural responses using various latent dynamical systems, our work demonstrates feasibility of fitting neural responses with low-dimensional recurrent circuits.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "low-dimensional recurrent circuits",
+      "label": "NeuralCircuit",
+      "sentence": "Although previous studies extensively modeled neural responses using various latent dynamical systems, our work demonstrates feasibility of fitting neural responses with low-dimensional recurrent circuits.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "neural data",
+      "label": "Measurement",
+      "sentence": "Most methods fitting low-dimensional dynamical systems to neural data do not explicitly model task inputs and behavior and do not ground the inferred dynamics in the underlying network connectivity32,33,40,52,53.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "RNN architectures",
+      "label": "ComputationalModel",
+      "sentence": "On the other hand, high-dimensional RNN architectures serve as intermediate mechanistic models of neural responses, which generate latent dynamics from high-dimensional recurrent connectivity25–29.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "On the other hand, high-dimensional RNN architectures serve as intermediate mechanistic models of neural responses, which generate latent dynamics from high-dimensional recurrent connectivity25–29.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "latent dynamics",
+      "label": "DynamicalSystemFeature",
+      "sentence": "On the other hand, high-dimensional RNN architectures serve as intermediate mechanistic models of neural responses, which generate latent dynamics from high-dimensional recurrent connectivity25–29.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "single-trial neural activity",
+      "label": "Measurement",
+      "sentence": "Our results motivate future work incorporating low-dimensional recurrent circuits as latent dynamics generators within these model architectures, which will enable combining their ability of fitting single-trial neural activity26,29 with the uniqueness and interpretability furnished by the latent circuit model.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "By relating connectivity to neural dynamics and behavior, the latent circuit model extends the causal predictive power of classical neural circuit models to the study of mechanisms of cognitive functions in high-dimensional networks.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "neural dynamics",
+      "label": "Phenomenon",
+      "sentence": "By relating connectivity to neural dynamics and behavior, the latent circuit model extends the causal predictive power of classical neural circuit models to the study of mechanisms of cognitive functions in high-dimensional networks.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Although dynamical mechanisms have been studied in RNNs by linearizing the RNN flow field around fixed points19,31, these dynamical mechanisms do not specify how a particular fixed-point configuration arises from the RNN connectivity.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "fixed points",
+      "label": "DynamicalSystemFeature",
+      "sentence": "Although dynamical mechanisms have been studied in RNNs by linearizing the RNN flow field around fixed points19,31, these dynamical mechanisms do not specify how a particular fixed-point configuration arises from the RNN connectivity.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "Although dynamical mechanisms have been studied in RNNs by linearizing the RNN flow field around fixed points19,31, these dynamical mechanisms do not specify how a particular fixed-point configuration arises from the RNN connectivity.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalVariable",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalVariable",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The inhibitory suppression mechanism revealed by the latent circuit model is qualitatively similar to previous neural circuit models of how PFC flexibly switches between alternative stimulus–response mappings8–10,44,45.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "inhibitory suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "The inhibitory suppression mechanism revealed by the latent circuit model is qualitatively similar to previous neural circuit models of how PFC flexibly switches between alternative stimulus–response mappings8–10,44,45.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "The inhibitory suppression mechanism revealed by the latent circuit model is qualitatively similar to previous neural circuit models of how PFC flexibly switches between alternative stimulus–response mappings8–10,44,45.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "stimulus–response mappings",
+      "label": "TaskParadigm",
+      "sentence": "The inhibitory suppression mechanism revealed by the latent circuit model is qualitatively similar to previous neural circuit models of how PFC flexibly switches between alternative stimulus–response mappings8–10,44,45.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "Thus, RNNs do not find qualitatively distinct solutions to this task, and complex selectivity of single neurons has a simple explanation as a linear mixing of the low-dimensional latent inhibitory circuit mechanism.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Thus, RNNs do not find qualitatively distinct solutions to this task, and complex selectivity of single neurons has a simple explanation as a linear mixing of the low-dimensional latent inhibitory circuit mechanism.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "correlation-based dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "By contrast, correlation-based dimensionality reduction methods found no significant suppression of irrelevant stimuli in the same data19,21.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We show that representations of irrelevant stimuli also exist in RNNs that provably implement an inhibitory suppression mechanism, but these representations do not causally drive choices.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "inhibitory suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "We show that representations of irrelevant stimuli also exist in RNNs that provably implement an inhibitory suppression mechanism, but these representations do not causally drive choices.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "cognitive flexibility",
+      "label": "Phenomenon",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit model opens a route for interpreting circuit mechanisms in high-dimensional networks.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "low-rank connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "Whereas low-dimensional data do not uniquely constrain the full high-dimensional connectivity in RNNs, we show that we can uniquely recover the latent low-rank connectivity within a high-dimensional network, which therefore can be reliably interpreted.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Whereas low-dimensional data do not uniquely constrain the full high-dimensional connectivity in RNNs, we show that we can uniquely recover the latent low-rank connectivity within a high-dimensional network, which therefore can be reliably interpreted.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "entity": "stomatogastric nervous system",
+      "label": "NervousSystemPart",
+      "sentence": "Marder, E. & Bucher, D. Understanding circuit dynamics using the stomatogastric nervous system of lobsters and crabs.",
+      "paper_location": "References; Page 11"
+    },
+    {
+      "entity": "lobsters",
+      "label": "Species",
+      "sentence": "Marder, E. & Bucher, D. Understanding circuit dynamics using the stomatogastric nervous system of lobsters and crabs.",
+      "paper_location": "References; Page 11"
+    },
+    {
+      "entity": "crabs",
+      "label": "Species",
+      "sentence": "Marder, E. & Bucher, D. Understanding circuit dynamics using the stomatogastric nervous system of lobsters and crabs.",
+      "paper_location": "References; Page 11"
+    },
+    {
+      "entity": "whisking oscillator circuit",
+      "label": "NeuralCircuit",
+      "sentence": "Takatoh, J. et al. The whisking oscillator circuit.",
+      "paper_location": "References; Page 11"
+    },
+    {
+      "entity": "mice",
+      "label": "Species",
+      "sentence": "Schaeffer, R., Khona, M., Meshulam, L., Brain Laboratory International & Fiete, I. Reverse-engineering recurrent neural network solutions to a hierarchical inference task for mice.",
+      "paper_location": "References; Page 11"
+    },
+    {
+      "entity": "schizophrenia",
+      "label": "Disease",
+      "sentence": "Murray, J. D. et al. Linking microcircuit dysfunction to cognitive impairment: effects of disinhibition associated with schizophrenia in a cortical working memory model.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "microcircuit dysfunction",
+      "label": "Phenotype",
+      "sentence": "Murray, J. D. et al. Linking microcircuit dysfunction to cognitive impairment: effects of disinhibition associated with schizophrenia in a cortical working memory model.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "cognitive impairment",
+      "label": "Phenotype",
+      "sentence": "Murray, J. D. et al. Linking microcircuit dysfunction to cognitive impairment: effects of disinhibition associated with schizophrenia in a cortical working memory model.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "disinhibition",
+      "label": "Phenomenon",
+      "sentence": "Murray, J. D. et al. Linking microcircuit dysfunction to cognitive impairment: effects of disinhibition associated with schizophrenia in a cortical working memory model.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "cortical working memory model",
+      "label": "ComputationalModel",
+      "sentence": "Murray, J. D. et al. Linking microcircuit dysfunction to cognitive impairment: effects of disinhibition associated with schizophrenia in a cortical working memory model.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "frontal cortex",
+      "label": "BrainRegion",
+      "sentence": "Inagaki, H. K., Fontolan, L., Romani, S. & Svoboda, K. Discrete attractor dynamics underlies persistent activity in the frontal cortex.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "persistent activity",
+      "label": "Phenomenon",
+      "sentence": "Inagaki, H. K., Fontolan, L., Romani, S. & Svoboda, K. Discrete attractor dynamics underlies persistent activity in the frontal cortex.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "discrete attractor dynamics",
+      "label": "DynamicalSystemFeature",
+      "sentence": "Inagaki, H. K., Fontolan, L., Romani, S. & Svoboda, K. Discrete attractor dynamics underlies persistent activity in the frontal cortex.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "Cromer, J. A., Roy, J. E. & Miller, E. K. Representation of multiple, independent categories in the primate prefrontal cortex.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "primate",
+      "label": "Species",
+      "sentence": "Cromer, J. A., Roy, J. E. & Miller, E. K. Representation of multiple, independent categories in the primate prefrontal cortex.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "mixed selectivity",
+      "label": "Phenomenon",
+      "sentence": "Rigotti, M. et al. The importance of mixed selectivity in complex cognitive tasks.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "Mante, V., Sussillo, D., Shenoy, K. V. & Newsome, W. T. Context-dependent computation by recurrent dynamics in prefrontal cortex.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "recurrent dynamics",
+      "label": "DynamicalSystemFeature",
+      "sentence": "Mante, V., Sussillo, D., Shenoy, K. V. & Newsome, W. T. Context-dependent computation by recurrent dynamics in prefrontal cortex.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "neural population",
+      "label": "CellType",
+      "sentence": "Raposo, D., Kaufman, M. T. & Churchland, A. K. A category-free neural population supports evolving demands during decision-making.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "Aoi, M. C., Mante, V. & Pillow, J. W. Prefrontal cortex exhibits multidimensional dynamic encoding during decision-making.",
+      "paper_location": "References; Page 10"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We fit the latent circuit model Eqs. (1)–(3) to neural response data y by minimizing the mean squared error loss function, L = ∑k∑t∥ ytk − Qxtk∥2+ ∥ ztk − woutxtk∥2, (7) using custom Python code59.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "We fit the latent circuit model Eqs. (1)–(3) to neural response data y by minimizing the mean squared error loss function, L = ∑k∑t∥ ytk − Qxtk∥2+ ∥ ztk − woutxtk∥2, (7) using custom Python code59.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "mean squared error loss function",
+      "label": "StatisticalMethod",
+      "sentence": "We fit the latent circuit model Eqs. (1)–(3) to neural response data y by minimizing the mean squared error loss function, L = ∑k∑t∥ ytk − Qxtk∥2+ ∥ ztk − woutxtk∥2, (7) using custom Python code59.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "We fit the latent circuit model Eqs. (1)–(3) to neural response data y by minimizing the mean squared error loss function, L = ∑k∑t∥ ytk − Qxtk∥2+ ∥ ztk − woutxtk∥2, (7) using custom Python code59.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "nonlinear least squares optimization",
+      "label": "StatisticalMethod",
+      "sentence": "Because the variable x depends implicitly on the latent circuit parameters wrec and win, the minimization of L is a nonlinear least squares optimization problem60 in which we simultaneously search for a behaviorally relevant projection of the high-dimensional activity and a low-dimensional neural circuit that generates dynamics in this projection.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "Cayley transform",
+      "label": "MathematicalMethod",
+      "sentence": "To transform it into an unconstrained problem, we use the Cayley transform to parameterize orthonormal matrices by the linear space of skew symmetric matrices61, Q = (I + A)(I − A)−1πn, (8)",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "We perform this minimization using PyTorch and the Adam optimizer with default values 0.9 and 0.999 for the decay rate of the first and second moment estimates, respectively, a learning rate of 0.02 and a weight decay of 0.001.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "Adam optimizer",
+      "label": "Software",
+      "sentence": "We perform this minimization using PyTorch and the Adam optimizer with default values 0.9 and 0.999 for the decay rate of the first and second moment estimates, respectively, a learning rate of 0.02 and a weight decay of 0.001.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "Seaborn",
+      "label": "Software",
+      "sentence": "We used the Python software package Seaborn for visualizing model parameters and responses after training.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "We used the Python software package Seaborn for visualizing model parameters and responses after training.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "permutation test",
+      "label": "StatisticalTest",
+      "sentence": "To determine whether the inferred latent circuit connectivity significantly depends on neural response data beyond the constraints imposed by the task alone, we performed a permutation test (Extended Data Figs. 1, 7c and 8c), which proceeds in three steps.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "To determine whether the inferred latent circuit connectivity significantly depends on neural response data beyond the constraints imposed by the task alone, we performed a permutation test (Extended Data Figs. 1, 7c and 8c), which proceeds in three steps.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "Mann–Whitney U-test",
+      "label": "StatisticalTest",
+      "sentence": "Third, we use a Mann–Whitney U-test to determine whether the correlation coefficients are significantly smaller for models fitted to the shuffled responses than original neural responses.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We consider RNNs of the form τẏ= −y + [Wrec y + Winu]+.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "rectified linear",
+      "label": "ActivationFunction",
+      "sentence": "Here, [⋅]+ is a rectified linear (ReLU) activation function, τ is a time constant, and u are external task inputs.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "ReLU",
+      "label": "ActivationFunction",
+      "sentence": "Here, [⋅]+ is a rectified linear (ReLU) activation function, τ is a time constant, and u are external task inputs.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "entity": "Softplus",
+      "label": "ActivationFunction",
+      "sentence": "To test whether our results extend to networks with a different biologically plausible nonlinearity, we trained RNNs that had a Softplus activation function f(x) = gβ log(1 + eβx) for a range of parameter β and also with varying gain g across units.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "To test whether our results extend to networks with a different biologically plausible nonlinearity, we trained RNNs that had a Softplus activation function f(x) = gβ log(1 + eβx) for a range of parameter β and also with varying gain g across units.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "RNN simulations",
+      "label": "Method",
+      "sentence": "RNN simulations",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "time-discretized RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We simulate dynamics of time-discretized RNNs using the general framework for modeling cognitive tasks22.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "recurrent units",
+      "label": "ComputationalUnit",
+      "sentence": "We consider RNNs with positive activity and N = 50 recurrent units.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "Euler scheme",
+      "label": "NumericalMethod",
+      "sentence": "We discretize the RNN dynamics Eq. (10) using the first-order Euler scheme with a time step Δt and add a noise term to obtain yt = (1 − α)yt−1 + α[Wrec yt−1 + Winut + √2ασrecξt]+.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "Dale’s law",
+      "label": "NeurobiologicalPrinciple",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale’s law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "excitatory units",
+      "label": "ComputationalUnit",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale’s law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "inhibitory units",
+      "label": "ComputationalUnit",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale’s law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "The RNN simulation and training were implemented in Python using the software package PyTorch.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "The RNN simulation and training were implemented in Python using the software package PyTorch.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "context cue",
+      "label": "Stimulus",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "sensory stimulus",
+      "label": "Stimulus",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "color coherence",
+      "label": "Measurement",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "context cues",
+      "label": "Stimulus",
+      "sentence": "To model the context-dependent decision-making task, the network receives six inputs u corresponding to two context cues (um: motion context; uc: color context) and sensory evidence streams for motion (um,L: motion left; um,R: motion right) and color (uc,R: color red; uc,G: color green).",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "sensory evidence streams",
+      "label": "Stimulus",
+      "sentence": "To model the context-dependent decision-making task, the network receives six inputs u corresponding to two context cues (um: motion context; uc: color context) and sensory evidence streams for motion (um,L: motion left; um,R: motion right) and color (uc,R: color red; uc,G: color green).",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "entity": "RNN training",
+      "label": "Method",
+      "sentence": "RNN training",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "firing rates",
+      "label": "Measurement",
+      "sentence": "The first term is the task error, and the second term serves to regularize by penalizing the magnitude of the firing rates.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "Adam algorithm",
+      "label": "Software",
+      "sentence": "The training is performed with the Adam algorithm.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "Dale’s law",
+      "label": "NeurobiologicalPrinciple",
+      "sentence": "To implement Dale’s law, connections are clipped to 0 after each training step if they change sign.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "Linear decoding",
+      "label": "Method",
+      "sentence": "Linear decoding",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "linear regression model",
+      "label": "StatisticalMethod",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model c = βy + b, (27) where β ∈ ℝ1×N is the vector of regression coefficients, c ∈ ℝ1×K⋅T is the motion coherence on each trial, b ∈ ℝ is a bias term, and y ∈ ℝN×K⋅T is the RNN responses at each time step during the stimulus epoch of each trial.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model c = βy + b, (27) where β ∈ ℝ1×N is the vector of regression coefficients, c ∈ ℝ1×K⋅T is the motion coherence on each trial, b ∈ ℝ is a bias term, and y ∈ ℝN×K⋅T is the RNN responses at each time step during the stimulus epoch of each trial.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model c = βy + b, (27) where β ∈ ℝ1×N is the vector of regression coefficients, c ∈ ℝ1×K⋅T is the motion coherence on each trial, b ∈ ℝ is a bias term, and y ∈ ℝN×K⋅T is the RNN responses at each time step during the stimulus epoch of each trial.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "frontal eye field",
+      "label": "BrainRegion",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "neural activity recordings",
+      "label": "Method",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "PFC data",
+      "label": "Measurement",
+      "sentence": "We fitted latent circuit models to the PFC data following similar procedures as for RNNs.",
+      "paper_location": "Methods; Page 15"
+    },
+    {
+      "entity": "latent circuit models",
+      "label": "ComputationalModel",
+      "sentence": "We fitted latent circuit models to the PFC data following similar procedures as for RNNs.",
+      "paper_location": "Methods; Page 15"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We fitted latent circuit models to the PFC data following similar procedures as for RNNs.",
+      "paper_location": "Methods; Page 15"
+    },
+    {
+      "entity": "saccade",
+      "label": "Behavior",
+      "sentence": "Monkeys reported their choice with a saccade to one of two targets presented shortly after fixation for the entire trial duration.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "fixation",
+      "label": "Behavior",
+      "sentence": "Monkeys reported their choice with a saccade to one of two targets presented shortly after fixation for the entire trial duration.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "entity": "spikes",
+      "label": "Measurement",
+      "sentence": "For each trial, we computed time-varying firing rates by counting spikes in a 50-ms sliding square window (50-ms steps).",
+      "paper_location": "Methods; Page 15"
+    },
+    {
+      "entity": "firing rates",
+      "label": "Measurement",
+      "sentence": "For each trial, we computed time-varying firing rates by counting spikes in a 50-ms sliding square window (50-ms steps).",
+      "paper_location": "Methods; Page 15"
+    },
+    {
+      "entity": "Gaussian kernel",
+      "label": "StatisticalMethod",
+      "sentence": "Within the training and test sets, we z scored and smoothed (Gaussian kernel, σ = 40 ms) the response of each unit.",
+      "paper_location": "Methods; Page 15"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "Last, to denoise these trajectories, we projected them onto the principal components explaining 50% of their total variance (corresponding to the first 40 and 31 principal components for monkeys A and F, respectively).",
+      "paper_location": "Methods; Page 15"
+    },
+    {
+      "entity": "National Institutes of Health",
+      "label": "Organization",
+      "sentence": "This work was supported by National Institutes of Health grants RF1DA055666 and S10OD028632-01 (T.A.E.), the Alfred P. Sloan Foundation Research Fellowship (T.A.E.) and the Swartz Foundation Postdoctoral Fellowship (C.L.).",
+      "paper_location": "Acknowledgements; Page 15"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "Recurrent neural networks were trained using PyTorch (version 2.4).",
+      "paper_location": "Reporting Summary; Page 27"
+    },
+    {
+      "entity": "Seaborn",
+      "label": "Software",
+      "sentence": "Plots were generated with Python (version 3.8) and Seaborn (version 0.9.0).",
+      "paper_location": "Reporting Summary; Page 27"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "Plots were generated with Python (version 3.8) and Seaborn (version 0.9.0).",
+      "paper_location": "Reporting Summary; Page 27"
+    },
+    {
+      "entity": "Plexon",
+      "label": "Instrument",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary; Page 27"
+    },
+    {
+      "entity": "MAP data-acquisition system",
+      "label": "Instrument",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary; Page 27"
+    },
+    {
+      "entity": "principal component analysis",
+      "label": "Method",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary; Page 27"
+    },
+    {
+      "entity": "Plexon offline sorter",
+      "label": "Software",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary; Page 27"
+    },
+    {
+      "entity": "Spike sorting",
+      "label": "Method",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary; Page 27"
+    },
+    {
+      "entity": "macaque monkeys",
+      "label": "Species",
+      "sentence": "Findings were successfully replicated in two macaque monkeys.",
+      "paper_location": "Reporting Summary; Page 28"
+    },
+    {
+      "entity": "adult males",
+      "label": "Species",
+      "sentence": "The animals were adult males, ages 13 (monkey A) and 8 (monkey F) years old at the time the experiments were conducted.",
+      "paper_location": "Reporting Summary; Page 29"
+    },
+    {
+      "entity": "National Institutes of Health",
+      "label": "Organization",
+      "sentence": "All surgical and behavioral procedures conformed to the guidelines established by the National Institutes of Health and were approved by the Institutional Animal Care and Use Committee of Stanford University.",
+      "paper_location": "Reporting Summary; Page 29"
+    }
+  ],
+  "key_terms": [
+    {
+      "term": "latent circuit inference",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Page 1"
+    },
+    {
+      "term": "context-dependent decision-making task",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Page 1"
+    },
+    {
+      "term": "patterned connectivity perturbations",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Page 1"
+    },
+    {
+      "term": "causal interactions among task variables",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Page 1"
+    },
+    {
+      "term": "hand-crafted neural circuit models",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Page 1"
+    },
+    {
+      "term": "mixed selectivity",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Page 3"
+    },
+    {
+      "term": "correlation-based dimensionality reduction methods",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC’s role in context-dependent decision-making.",
+      "paper_location": "Page 2"
+    },
+    {
+      "term": "stimulus–response mappings",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus–response mappings.",
+      "paper_location": "Page 2"
+    },
+    {
+      "term": "low-dimensional circuit mechanism",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Page 2"
+    },
+    {
+      "term": "behaviorally irrelevant representations",
+      "sentence": "Using RNN perturbations, we show that dimensionality reduction methods that do not incorporate causal interactions among latent variables are biased toward uncovering behaviorally irrelevant representations.",
+      "paper_location": "Page 2"
+    },
+    {
+      "term": "orthonormal embedding matrix",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as y = Qx, (1) where Q ∈ ℝN×n is an orthonormal embedding matrix and n ≪ N.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "term": "rectified linear (ReLU) activation function",
+      "sentence": "The latent variables x are constrained to be nodes in a neural circuit with dynamics ẋ= −x + f(wrecx + winu), (2) where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "term": "low-dimensional subspace",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. (2).",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "term": "orthonormality constraint",
+      "sentence": "The orthonormality constraint on Q implies that the projection defined by the transpose matrix QT is a dimensionality reduction in which projection onto the ith column of Q correlates with the activity of the ith node in the latent circuit.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "term": "low-rank structure",
+      "sentence": "The relation Eq. (5) shows that the latent circuit connectivity wrec is a low-rank structure in the connectivity of a high-dimensional network, which captures interactions among the latent variables defined by the columns of Q.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "term": "invariant subspace",
+      "sentence": "Rather, it is a weaker condition that the linear subspace defined by Q is an invariant subspace of the high-dimensional recurrent connectivity matrix.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "term": "rank-one perturbation",
+      "sentence": "Specifically, a change in the connection δij between nodes i and j in the latent circuit maps onto a rank-one perturbation of the RNN connectivity matrix (Fig. 1e and Methods), δij → qiqTj, (6) where qi is the ith column of Q.",
+      "paper_location": "Results; Page 3"
+    },
+    {
+      "term": "color context",
+      "sentence": "In the color context, the choice should be made according to the color and ignoring the motion stimulus and vice versa in the motion context.",
+      "paper_location": "Figure 2; Page 4"
+    },
+    {
+      "term": "motion context",
+      "sentence": "In the color context, the choice should be made according to the color and ignoring the motion stimulus and vice versa in the motion context.",
+      "paper_location": "Figure 2; Page 4"
+    },
+    {
+      "term": "psychometric functions",
+      "sentence": "Psychometric functions show that the RNN successfully learns the task; it responds to relevant stimuli and ignores irrelevant stimuli in each context.",
+      "paper_location": "Figure 2; Page 4"
+    },
+    {
+      "term": "task variables",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "term": "inhibition of irrelevant representations",
+      "sentence": "This suppression mechanism based on inhibition of irrelevant representations is qualitatively similar to mechanisms for context-dependent decision-making hypothesized in previous hand-crafted neural circuit models8,9.",
+      "paper_location": "Results; Page 4"
+    },
+    {
+      "term": "suppression of irrelevant sensory representations",
+      "sentence": "The suppression of irrelevant sensory representations in RNN activity is consistent with the inhibitory mechanism revealed in the latent circuit connectivity wrec.",
+      "paper_location": "Results; Page 5"
+    },
+    {
+      "term": "latent circuit solutions",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices wrec and projected the data onto the first two principal components, which accounted for 42% of total variance (Fig. 5a).",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "term": "solution clusters",
+      "sentence": "The latent circuit solutions from all RNNs formed three major clusters.",
+      "paper_location": "Results; Page 6"
+    },
+    {
+      "term": "inhibitory mechanism",
+      "sentence": "Although the circuit solutions in two of the clusters exploit asymmetries in the representations of sensory evidence, they still operate by an inhibitory mechanism in which irrelevant responses are suppressed (Extended Data Fig. 3).",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "term": "dynamical selection vector mechanism",
+      "sentence": "We analyzed dynamics in our RNNs and found the same selection vector mechanism (Extended Data Fig. 4), indicating that the dynamical selection vector mechanism is a local linear description of the inhibitory circuit mechanism.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "term": "task subspace",
+      "sentence": "The task subspace Q explained a smaller fraction of total variance in the PFC responses (11.0% and 7.0% for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC42 and reflects the high dimensionality of PFC activity19.",
+      "paper_location": "Results; Page 7"
+    },
+    {
+      "term": "error trials",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "term": "incongruent trials",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "term": "behavioral relevance",
+      "sentence": "To further test for behavioral relevance of representations identified by different dimensionality reduction methods, we again turned to RNNs in which we can measure behavioral effects of arbitrary activity perturbations.",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "term": "decoder axis",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results; Page 8"
+    },
+    {
+      "term": "decision boundary",
+      "sentence": "As expected, driving RNN activity along the motion axis of the latent circuit model shifted the decision boundary on motion context trials (Fig. 7d).",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "term": "demixing representations",
+      "sentence": "Our results indicate that ‘demixing’ representations of task variables19,21,42 may not be the right objective for identifying behaviorally relevant patterns in neural activity and may provide a misleading picture of computation.",
+      "paper_location": "Results; Page 9"
+    },
+    {
+      "term": "single-trial neural activity",
+      "sentence": "Our results motivate future work incorporating low-dimensional recurrent circuits as latent dynamics generators within these model architectures, which will enable combining their ability of fitting single-trial neural activity26,29 with the uniqueness and interpretability furnished by the latent circuit model.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "term": "causal predictive power",
+      "sentence": "By relating connectivity to neural dynamics and behavior, the latent circuit model extends the causal predictive power of classical neural circuit models to the study of mechanisms of cognitive functions in high-dimensional networks.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "term": "out-of-distribution inputs",
+      "sentence": "The causal predictive power of the latent circuit model is further supported by its ability to predict RNN dynamics and behavior for out-of-distribution inputs (Supplementary Fig. 8).",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "term": "low-rank RNNs",
+      "sentence": "However, it was unclear whether low-rank RNNs use mechanisms similar to classical circuit models or implement truly novel solutions that emerge only in high-dimensional nonlinear systems.",
+      "paper_location": "Discussion; Page 9"
+    },
+    {
+      "term": "cognitive flexibility",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "term": "circuit dissection methods",
+      "sentence": "Our work translates such circuit dissection methods56,57 to high-dimensional networks in which, unlike in small circuits, individual connections do not carry specific functions, but instead these functions arise from distributed connectivity patterns.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "term": "cognitive control tasks",
+      "sentence": "Consistent with this idea, we find that RNNs trained on different tasks use a similar inhibitory control mechanism (Supplementary Fig. 7), which may generalize to other cognitive control tasks as well.",
+      "paper_location": "Discussion; Page 10"
+    },
+    {
+      "term": "nonlinear least squares optimization",
+      "sentence": "Because the variable x depends implicitly on the latent circuit parameters wrec and win, the minimization of L is a nonlinear least squares optimization problem60 in which we simultaneously search for a behaviorally relevant projection of the high-dimensional activity and a low-dimensional neural circuit that generates dynamics in this projection.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "term": "constrained optimization problem",
+      "sentence": "Because orthonormal matrices define a nonlinear submanifold within the space of all matrices, minimizing L corresponds to solving a constrained optimization problem over this submanifold.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "term": "minibatch size",
+      "sentence": "We use a minibatch size of 128 trials.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "term": "held out test data",
+      "sentence": "Therefore, we selected the best ten latent circuit models from this ensemble in terms of fit quality on held out test data, which formed a set of converged solutions (Fig. 5a).",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "term": "shuffling procedure",
+      "sentence": "Our shuffling procedure randomly permutes neural responses with respect to trial conditions while preserving the input–output relationship on each trial so that the fitted latent circuit models can still perform the task.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "term": "input–output relationship",
+      "sentence": "Our shuffling procedure randomly permutes neural responses with respect to trial conditions while preserving the input–output relationship on each trial so that the fitted latent circuit models can still perform the task.",
+      "paper_location": "Methods; Page 12"
+    },
+    {
+      "term": "piecewise-linear systems",
+      "sentence": "Because this is an equality of two piecewise-linear systems, it holds for each local linear piece individually.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "term": "Softplus activation function",
+      "sentence": "To test whether our results extend to networks with a different biologically plausible nonlinearity, we trained RNNs that had a Softplus activation function f(x) = gβ log(1 + eβx) for a range of parameter β and also with varying gain g across units.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "term": "first-order Euler scheme",
+      "sentence": "We discretize the RNN dynamics Eq. (10) using the first-order Euler scheme with a time step Δt and add a noise term to obtain yt = (1 − α)yt−1 + α[Wrec yt−1 + Winut + √2ασrecξt]+.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "term": "Dale’s law",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale’s law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "term": "motion and color sensory evidence",
+      "sentence": "After a delay of 200 ms, so that the network must maintain a memory of the context cue, the inputs corresponding to motion and color sensory evidence are presented at t = 1,200 ms for the remaining duration of the trial.",
+      "paper_location": "Methods; Page 13"
+    },
+    {
+      "term": "input noise strength",
+      "sentence": "The input noise strength is σin = 0.01.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "term": "mean squared error",
+      "sentence": "To train the RNN, we minimize the mean squared error between the output z(t) and the target ztarget(t): ℒ ∶= ∑ikt(zikt − ztarget,ikt)2 + λr ∑ikty2ikt.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "term": "L2 penalty",
+      "sentence": "We control the degree of correlation between the input and output vectors in the RNN by adding an L2 penalty λorth ∥ BTB − diag (BTB)∥2 (26) to the loss function in Eq. (25) during training.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "term": "linear regression model",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model c = βy + b, (27) where β ∈ ℝ1×N is the vector of regression coefficients, c ∈ ℝ1×K⋅T is the motion coherence on each trial, b ∈ ℝ is a bias term, and y ∈ ℝN×K⋅T is the RNN responses at each time step during the stimulus epoch of each trial.",
+      "paper_location": "Methods; Page 14"
+    },
+    {
+      "term": "neurophysiological data",
+      "sentence": "The neurophysiological data were previously described in Ref. 19 (Mante et al., Nature, 503, 78, 2013).",
+      "paper_location": "Reporting Summary; Page 27"
+    },
+    {
+      "term": "data-acquisition system",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary; Page 27"
+    },
+    {
+      "term": "spike sorting",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary; Page 27"
+    },
+    {
+      "term": "source code",
+      "sentence": "The source code to reproduce results of this study is available on GitHub (https://github.com/engellab/latentcircuit) and Zenodo (https://zenodo.org/records/14020108).",
+      "paper_location": "Reporting Summary; Page 27"
+    },
+    {
+      "term": "public repository",
+      "sentence": "The neurophysiological data were previously described in Ref. 19 (Mante et al., Nature, 503, 78, 2013) and are available in the public repository https://www.ini.uzh.ch/en/research/groups/mante/data.html for download.",
+      "paper_location": "Reporting Summary; Page 28"
+    },
+    {
+      "term": "surgical and behavioral procedures",
+      "sentence": "All surgical and behavioral procedures conformed to the guidelines established by the National Institutes of Health and were approved by the Institutional Animal Care and Use Committee of Stanford University.",
+      "paper_location": "Reporting Summary; Page 29"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_gpt-5.5_20260629_085630.json b/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_gpt-5.5_20260629_085630.json
new file mode 100644
index 0000000..8e85275
--- /dev/null
+++ b/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_gpt-5.5_20260629_085630.json
@@ -0,0 +1,2710 @@
+{
+  "source_metadata": {
+    "source_path": "evaluation/ner/latent-circuit-inference/s41593-025-01869-7.pdf"
+  },
+  "entities": [
+    {
+      "entity": "Higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "Dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "dimensionality reduction",
+      "label": "Method",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "recurrent connectivity",
+      "label": "Measurement",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "recurrent neural networks",
+      "label": "ComputationalModel",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "contextual representations",
+      "label": "NeuralRepresentation",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "sensory responses",
+      "label": "Measurement",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "sensory responses",
+      "label": "Measurement",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "neural response data",
+      "label": "DataType",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "Cognitive functions",
+      "label": "Phenomenon",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "sensory",
+      "label": "Stimulus",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "contextual signals",
+      "label": "Stimulus",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral response",
+      "label": "Measurement",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "excitatory",
+      "label": "CellType",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "inhibitory neurons",
+      "label": "CellType",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cortical circuits",
+      "label": "NeuralCircuit",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural populations",
+      "label": "CellType",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "circuit models",
+      "label": "ComputationalModel",
+      "sentence": "Because these circuit models usually assume a relatively simple connectivity structure, their connectivity can be directly related to a dynamical system description of computations supporting cognitive task execution5,6.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "dynamical system",
+      "label": "ComputationalModel",
+      "sentence": "Because these circuit models usually assume a relatively simple connectivity structure, their connectivity can be directly related to a dynamical system description of computations supporting cognitive task execution5,6.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cognitive task",
+      "label": "BehavioralAssay",
+      "sentence": "Because these circuit models usually assume a relatively simple connectivity structure, their connectivity can be directly related to a dynamical system description of computations supporting cognitive task execution5,6.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural dynamics",
+      "label": "Phenomenon",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation–inhibition balance12) and thus can be causally validated in experiments13–15.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation–inhibition balance12) and thus can be causally validated in experiments13–15.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral performance",
+      "label": "Measurement",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation–inhibition balance12) and thus can be causally validated in experiments13–15.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "excitation–inhibition balance",
+      "label": "Phenomenon",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation–inhibition balance12) and thus can be causally validated in experiments13–15.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of single-neuron responses in the cortex.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cortex",
+      "label": "BrainRegion",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of single-neuron responses in the cortex.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "Single neurons",
+      "label": "CellType",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "heterogeneous tuning",
+      "label": "Measurement",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "recurrent neural network",
+      "label": "ComputationalModel",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural activity data",
+      "label": "DataType",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "recurrent connectivity parameters",
+      "label": "Measurement",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks30.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks30.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "high-dimensional activity",
+      "label": "Measurement",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks30.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "dynamical system",
+      "label": "ComputationalModel",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural response data",
+      "label": "DataType",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "high-dimensional RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Thus, high-dimensional RNNs currently serve as intermediate mechanistic models of heterogeneous neural responses, which yield dynamical mechanisms but leave the underlying circuit mechanisms unknown.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "heterogeneous neural responses",
+      "label": "Measurement",
+      "sentence": "Thus, high-dimensional RNNs currently serve as intermediate mechanistic models of heterogeneous neural responses, which yield dynamical mechanisms but leave the underlying circuit mechanisms unknown.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas30,36.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "population activity",
+      "label": "Measurement",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas30,36.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas30,36.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "brain areas",
+      "label": "BrainRegion",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas30,36.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural population activity",
+      "label": "Measurement",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "unsupervised methods",
+      "label": "Method",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution26,37–40.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral outputs",
+      "label": "Measurement",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution26,37–40.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cognitive task",
+      "label": "BehavioralAssay",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution26,37–40.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "targeted dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural representations",
+      "label": "NeuralRepresentation",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral outputs",
+      "label": "Measurement",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural population activity",
+      "label": "Measurement",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "However, unlike RNNs and circuit models, these correlation-based methods do not incorporate recurrent interactions among task variables, which implement computations necessary to solve the task.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "circuit models",
+      "label": "ComputationalModel",
+      "sentence": "However, unlike RNNs and circuit models, these correlation-based methods do not incorporate recurrent interactions among task variables, which implement computations necessary to solve the task.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC’s role in context-dependent decision-making.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC’s role in context-dependent decision-making.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "Context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus–response mappings.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "stimulus–response mappings",
+      "label": "Phenomenon",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus–response mappings.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "NeuralRepresentation",
+      "sentence": "Most circuit models hypothesize a relatively simple mechanism based on inhibition of sensory representations irrelevant in the current context8–10,44–46.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC data",
+      "label": "DataType",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "sensory responses",
+      "label": "Measurement",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "These results suggest that RNNs and conceivably the PFC implement qualitatively novel task solutions that do not exist in small circuits and emerge only in high-dimensional networks19,47,48.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "These results suggest that RNNs and conceivably the PFC implement qualitatively novel task solutions that do not exist in small circuits and emerge only in high-dimensional networks19,47,48.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "high-dimensional networks",
+      "label": "ComputationalModel",
+      "sentence": "These results suggest that RNNs and conceivably the PFC implement qualitatively novel task solutions that do not exist in small circuits and emerge only in high-dimensional networks19,47,48.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "dimensionality reduction",
+      "label": "Method",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "heterogeneous neural responses",
+      "label": "Measurement",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "NeuralRepresentation",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We validated this mechanism by confirming the behavioral effects of patterned perturbations of the RNN activity and connectivity predicted by the latent circuit model.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We validated this mechanism by confirming the behavioral effects of patterned perturbations of the RNN activity and connectivity predicted by the latent circuit model.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural recordings",
+      "label": "DataType",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNN perturbations",
+      "label": "Method",
+      "sentence": "Using RNN perturbations, we show that dimensionality reduction methods that do not incorporate causal interactions among latent variables are biased toward uncovering behaviorally irrelevant representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Using RNN perturbations, we show that dimensionality reduction methods that do not incorporate causal interactions among latent variables are biased toward uncovering behaviorally irrelevant representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "dimensionality reduction",
+      "label": "Method",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "single-neuron heterogeneity",
+      "label": "Measurement",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as y = Qx, (1) where Q ∈ ℝN×n is an orthonormal embedding matrix and n ≪ N.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "high-dimensional neural responses",
+      "label": "Measurement",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as y = Qx, (1) where Q ∈ ℝN×n is an orthonormal embedding matrix and n ≪ N.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neurons",
+      "label": "CellType",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as y = Qx, (1) where Q ∈ ℝN×n is an orthonormal embedding matrix and n ≪ N.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "cognitive task",
+      "label": "BehavioralAssay",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as y = Qx, (1) where Q ∈ ℝN×n is an orthonormal embedding matrix and n ≪ N.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neural circuit",
+      "label": "ComputationalModel",
+      "sentence": "The latent variables x are constrained to be nodes in a neural circuit with dynamics ẋ= −x + f(wrecx + winu), (2) where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "rectified linear",
+      "label": "ActivationFunction",
+      "sentence": "The latent variables x are constrained to be nodes in a neural circuit with dynamics ẋ= −x + f(wrecx + winu), (2) where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "ReLU",
+      "label": "ActivationFunction",
+      "sentence": "The latent variables x are constrained to be nodes in a neural circuit with dynamics ẋ= −x + f(wrecx + winu), (2) where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "The latent nodes interact via the recurrent connectivity wrec and receive external task inputs u through the input connectivity win.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "recurrent connectivity",
+      "label": "Measurement",
+      "sentence": "The latent nodes interact via the recurrent connectivity wrec and receive external task inputs u through the input connectivity win.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "input connectivity",
+      "label": "Measurement",
+      "sentence": "The latent nodes interact via the recurrent connectivity wrec and receive external task inputs u through the input connectivity win.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity wout: z = woutx.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "circuit activity",
+      "label": "Measurement",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity wout: z = woutx.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "output connectivity",
+      "label": "Measurement",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity wout: z = woutx.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. (2).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. (2).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neural circuit",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. (2).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent circuit parameters",
+      "label": "ComputationalModelParameter",
+      "sentence": "We infer the latent circuit parameters (Q, wrec, win and wout) from neural activity y by minimizing the loss function L = ∑k,t∥y − Qx∥2 + ∥z − woutx∥2, where k and t index trials and time within a trial, respectively (Methods).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "We infer the latent circuit parameters (Q, wrec, win and wout) from neural activity y by minimizing the loss function L = ∑k,t∥y − Qx∥2 + ∥z − woutx∥2, where k and t index trials and time within a trial, respectively (Methods).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "loss function",
+      "label": "StatisticalMethod",
+      "sentence": "We infer the latent circuit parameters (Q, wrec, win and wout) from neural activity y by minimizing the loss function L = ∑k,t∥y − Qx∥2 + ∥z − woutx∥2, where k and t index trials and time within a trial, respectively (Methods).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "cognitive task",
+      "label": "BehavioralAssay",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "mixed selectivity",
+      "label": "Phenomenon",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "cortex",
+      "label": "BrainRegion",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "network connectivity",
+      "label": "Measurement",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "behavioral outcomes",
+      "label": "Measurement",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We consider RNNs with dynamics ẏ= −y + f(Wrec y + Winu), (4) where Wrec and Win are the recurrent and input connectivity matrices, respectively (Methods).",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "recurrent",
+      "label": "Measurement",
+      "sentence": "We consider RNNs with dynamics ẏ= −y + f(Wrec y + Winu), (4) where Wrec and Win are the recurrent and input connectivity matrices, respectively (Methods).",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "input connectivity matrices",
+      "label": "Measurement",
+      "sentence": "We consider RNNs with dynamics ẏ= −y + f(Wrec y + Winu), (4) where Wrec and Win are the recurrent and input connectivity matrices, respectively (Methods).",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "The relation between connectivity matrices has the powerful consequence that we can validate the latent circuit mechanism directly in the RNN connectivity.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "Measurement",
+      "sentence": "The relation between connectivity matrices has the powerful consequence that we can validate the latent circuit mechanism directly in the RNN connectivity.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "The latent circuit inference uses only RNN activity without knowledge of the RNN connectivity.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "The latent circuit inference uses only RNN activity without knowledge of the RNN connectivity.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "Measurement",
+      "sentence": "The latent circuit inference uses only RNN activity without knowledge of the RNN connectivity.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory stimulus",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context cue",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "The RNN successfully learns the task; it makes choices according to the relevant stimulus and ignores the irrelevant stimulus in each context (Fig. 2c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "relevant stimulus",
+      "label": "Stimulus",
+      "sentence": "The RNN successfully learns the task; it makes choices according to the relevant stimulus and ignores the irrelevant stimulus in each context (Fig. 2c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant stimulus",
+      "label": "Stimulus",
+      "sentence": "The RNN successfully learns the task; it makes choices according to the relevant stimulus and ignores the irrelevant stimulus in each context (Fig. 2c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalModelComponent",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "mixed selectivity",
+      "label": "Phenomenon",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "Measurement",
+      "sentence": "The structure in the RNN connectivity responsible for generating the correct behavioral outputs is not immediately obvious (Fig. 2d).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "behavioral outputs",
+      "label": "Measurement",
+      "sentence": "The structure in the RNN connectivity responsible for generating the correct behavioral outputs is not immediately obvious (Fig. 2d).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We fitted the latent circuit model to the responses of RNN units and its output behavior during the task.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalModelComponent",
+      "sentence": "We fitted the latent circuit model to the responses of RNN units and its output behavior during the task.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory color nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory motion nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "choice nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "This choice of the latent circuit dimensionality agrees with the dimensionality of RNN responses after training, which is usually close to the total number of inputs and outputs (the first eight principal components accounted for 97.9% of the total variance in RNN responses; Supplementary Fig. 2).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "This choice of the latent circuit dimensionality agrees with the dimensionality of RNN responses after training, which is usually close to the total number of inputs and outputs (the first eight principal components accounted for 97.9% of the total variance in RNN responses; Supplementary Fig. 2).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "This choice of the latent circuit dimensionality agrees with the dimensionality of RNN responses after training, which is usually close to the total number of inputs and outputs (the first eight principal components accounted for 97.9% of the total variance in RNN responses; Supplementary Fig. 2).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "The fitted latent circuit model captured an overwhelming amount of variance in the RNN activity (coefficient of determination r2 = 0.96 on test data) and accurately matched projected RNN trajectories (Supplementary Fig. 3).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "The fitted latent circuit model captured an overwhelming amount of variance in the RNN activity (coefficient of determination r2 = 0.96 on test data) and accurately matched projected RNN trajectories (Supplementary Fig. 3).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "coefficient of determination",
+      "label": "StatisticalMeasure",
+      "sentence": "The fitted latent circuit model captured an overwhelming amount of variance in the RNN activity (coefficient of determination r2 = 0.96 on test data) and accurately matched projected RNN trajectories (Supplementary Fig. 3).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN trajectories",
+      "label": "Measurement",
+      "sentence": "The fitted latent circuit model captured an overwhelming amount of variance in the RNN activity (coefficient of determination r2 = 0.96 on test data) and accurately matched projected RNN trajectories (Supplementary Fig. 3).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "recurrent connectivity",
+      "label": "Measurement",
+      "sentence": "The inferred recurrent connectivity wrec of the latent circuit revealed an interpretable mechanism for context-dependent decision-making (Fig. 3a,b).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "The inferred recurrent connectivity wrec of the latent circuit revealed an interpretable mechanism for context-dependent decision-making (Fig. 3a,b).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "The inferred recurrent connectivity wrec of the latent circuit revealed an interpretable mechanism for context-dependent decision-making (Fig. 3a,b).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "stimuli",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "stimuli",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "stimulus–response mappings",
+      "label": "Phenomenon",
+      "sentence": "This pattern of connections from sensory to choice nodes implements two alternative stimulus–response mappings in the task.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color context node",
+      "label": "ComputationalModelComponent",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion context node",
+      "label": "ComputationalModelComponent",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "This pattern of connections from the context nodes to the sensory nodes implements a suppression mechanism that inhibits the irrelevant stimulus–response mapping in each context.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "This pattern of connections from the context nodes to the sensory nodes implements a suppression mechanism that inhibits the irrelevant stimulus–response mapping in each context.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "stimulus–response mapping",
+      "label": "Phenomenon",
+      "sentence": "This pattern of connections from the context nodes to the sensory nodes implements a suppression mechanism that inhibits the irrelevant stimulus–response mapping in each context.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory representation",
+      "label": "NeuralRepresentation",
+      "sentence": "Because the irrelevant sensory representation is suppressed, it does not drive the decision output.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "decision output",
+      "label": "Measurement",
+      "sentence": "Because the irrelevant sensory representation is suppressed, it does not drive the decision output.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "This suppression mechanism based on inhibition of irrelevant representations is qualitatively similar to mechanisms for context-dependent decision-making hypothesized in previous hand-crafted neural circuit models8,9.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "This suppression mechanism based on inhibition of irrelevant representations is qualitatively similar to mechanisms for context-dependent decision-making hypothesized in previous hand-crafted neural circuit models8,9.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "First, we verified the signatures of the suppression mechanism in the RNN activity.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "We projected RNN responses onto the columns of Q, which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity patterns",
+      "label": "Measurement",
+      "sentence": "We projected RNN responses onto the columns of Q, which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "We projected RNN responses onto the columns of Q, which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "choice axis",
+      "label": "NeuralRepresentation",
+      "sentence": "Next, the choice axis is the difference of two columns of Q corresponding to the left and right choice nodes in the latent circuit.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion axis",
+      "label": "NeuralRepresentation",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color axis",
+      "label": "NeuralRepresentation",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "Projections of RNN activity onto the motion and color axes reveal representations of relevant sensory stimuli, whereas representations of irrelevant stimuli are suppressed.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory stimuli",
+      "label": "Stimulus",
+      "sentence": "Projections of RNN activity onto the motion and color axes reveal representations of relevant sensory stimuli, whereas representations of irrelevant stimuli are suppressed.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "Measurement",
+      "sentence": "Second, we used the connectivity relationships Eq. (5) to directly validate the latent circuit mechanism in the RNN connectivity.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "embedding matrix",
+      "label": "ComputationalModelParameter",
+      "sentence": "We conjugated the RNN connectivity matrices with the embedding matrix Q.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "The resulting matrices closely match the connectivity in the latent circuit (Fig. 3d; correlation coefficient r = 0.89).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "correlation coefficient",
+      "label": "StatisticalMeasure",
+      "sentence": "The resulting matrices closely match the connectivity in the latent circuit (Fig. 3d; correlation coefficient r = 0.89).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "Measurement",
+      "sentence": "Finally, to ultimately validate that this latent connectivity structure supports the behavioral task performance, we tested whether patterned perturbations of the RNN connectivity (Eq. (6)) produced the same behavioral effects as predicted by the latent circuit model.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "Finally, to ultimately validate that this latent connectivity structure supports the behavioral task performance, we tested whether patterned perturbations of the RNN connectivity (Eq. (6)) produced the same behavioral effects as predicted by the latent circuit model.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "psychometric function",
+      "label": "Measurement",
+      "sentence": "Indeed, weakening these connections in the latent circuit produced the predicted behavioral effect in the psychometric function, visible as a rotation of the decision boundary on motion context trials (Fig. 4a).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "Measurement",
+      "sentence": "Perturbations of the RNN connectivity along the corresponding pattern produced similar behavioral effects (Fig. 4b), confirming that this connectivity pattern implements suppression of irrelevant sensory representations in the RNN.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "NeuralRepresentation",
+      "sentence": "Perturbations of the RNN connectivity along the corresponding pattern produced similar behavioral effects (Fig. 4b), confirming that this connectivity pattern implements suppression of irrelevant sensory representations in the RNN.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Perturbations of the RNN connectivity along the corresponding pattern produced similar behavioral effects (Fig. 4b), confirming that this connectivity pattern implements suppression of irrelevant sensory representations in the RNN.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We next asked whether RNNs trained on the same task arrive at different circuit solutions for context-dependent decision-making.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "We next asked whether RNNs trained on the same task arrive at different circuit solutions for context-dependent decision-making.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "The latent circuit inference enables us to determine whether different RNNs use the same circuit mechanism.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit inference enables us to determine whether different RNNs use the same circuit mechanism.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "principal component analysis",
+      "label": "Method",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices wrec and projected the data onto the first two principal components, which accounted for 42% of total variance (Fig. 5a).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices wrec and projected the data onto the first two principal components, which accounted for 42% of total variance (Fig. 5a).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "Gaussian mixture model",
+      "label": "StatisticalMethod",
+      "sentence": "The projection reveals three major solution clusters, which we partitioned by fitting data with a Gaussian mixture model with three components.",
+      "paper_location": "Figure 5 legend"
+    },
+    {
+      "entity": "Mann–Whitney U-test",
+      "label": "StatisticalMethod",
+      "sentence": "These latent models produced a significantly worse fit when the latent circuit and the target RNN were sampled from different clusters than from the same cluster (Extended Data Fig. 2m; one-sided Mann–Whitney U-test, U = 52, 593, P < 10−10), confirming differences in circuit mechanisms across clusters.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "These latent models produced a significantly worse fit when the latent circuit and the target RNN were sampled from different clusters than from the same cluster (Extended Data Fig. 2m; one-sided Mann–Whitney U-test, U = 52, 593, P < 10−10), confirming differences in circuit mechanisms across clusters.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "The mean connectivity matrices revealed circuits that all showed signatures of the suppression mechanism, with context nodes inhibiting irrelevant sensory nodes (Fig. 5b).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "The mean connectivity matrices revealed circuits that all showed signatures of the suppression mechanism, with context nodes inhibiting irrelevant sensory nodes (Fig. 5b).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "dynamical selection vector mechanism",
+      "label": "ComputationalModel",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses19.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses19.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "sensory responses",
+      "label": "Measurement",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses19.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "selection vector mechanism",
+      "label": "ComputationalModel",
+      "sentence": "We analyzed dynamics in our RNNs and found the same selection vector mechanism (Extended Data Fig. 4), indicating that the dynamical selection vector mechanism is a local linear description of the inhibitory circuit mechanism.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We analyzed dynamics in our RNNs and found the same selection vector mechanism (Extended Data Fig. 4), indicating that the dynamical selection vector mechanism is a local linear description of the inhibitory circuit mechanism.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "Softplus",
+      "label": "ActivationFunction",
+      "sentence": "We trained RNNs with the Softplus activation function of units to perform the context-dependent decision-making task, using different values of the parameter β across RNNs (25 RNNs for each value of β).",
+      "paper_location": "Extended Data Fig. 5 legend"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We trained RNNs with the Softplus activation function of units to perform the context-dependent decision-making task, using different values of the parameter β across RNNs (25 RNNs for each value of β).",
+      "paper_location": "Extended Data Fig. 5 legend"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We trained RNNs with the Softplus activation function of units to perform the context-dependent decision-making task, using different values of the parameter β across RNNs (25 RNNs for each value of β).",
+      "paper_location": "Extended Data Fig. 5 legend"
+    },
+    {
+      "entity": "ReLU",
+      "label": "ActivationFunction",
+      "sentence": "Example connectivity of a latent circuit model with ReLU non-linearity fitted to responses of an RNN with Softplus non-linearity for β = 10.",
+      "paper_location": "Extended Data Fig. 5 legend"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Example connectivity of a latent circuit model with ReLU non-linearity fitted to responses of an RNN with Softplus non-linearity for β = 10.",
+      "paper_location": "Extended Data Fig. 5 legend"
+    },
+    {
+      "entity": "Softplus",
+      "label": "ActivationFunction",
+      "sentence": "Example connectivity of a latent circuit model with ReLU non-linearity fitted to responses of an RNN with Softplus non-linearity for β = 10.",
+      "paper_location": "Extended Data Fig. 5 legend"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "sensory responses",
+      "label": "Measurement",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "This conclusion was derived using dimensionality reduction methods that fit neural responses with regression models19,21,41–43 to find low-dimensional projections that best correlate with task variables (Fig. 1b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "This conclusion was derived using dimensionality reduction methods that fit neural responses with regression models19,21,41–43 to find low-dimensional projections that best correlate with task variables (Fig. 1b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "regression models",
+      "label": "StatisticalMethod",
+      "sentence": "This conclusion was derived using dimensionality reduction methods that fit neural responses with regression models19,21,41–43 to find low-dimensional projections that best correlate with task variables (Fig. 1b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC recordings",
+      "label": "DataType",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC neurons",
+      "label": "CellType",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "rhesus monkeys",
+      "label": "Species",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "smoothed condition-averaged PFC responses",
+      "label": "Measurement",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting 100 ms after the stimulus onset19 (Methods).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "stimulus onset",
+      "label": "StimulusTime",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting 100 ms after the stimulus onset19 (Methods).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "The latent circuit model provided good fits of PFC responses projected onto the low-dimensional subspace spanned by the columns of the inferred embedding matrix Q for both monkeys (r2 = 0.88 and 0.76 on test data for monkeys A and F, respectively; Extended Data Figs. 7a and 8a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "The latent circuit model provided good fits of PFC responses projected onto the low-dimensional subspace spanned by the columns of the inferred embedding matrix Q for both monkeys (r2 = 0.88 and 0.76 on test data for monkeys A and F, respectively; Extended Data Figs. 7a and 8a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "The task subspace Q explained a smaller fraction of total variance in the PFC responses (11.0% and 7.0% for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC42 and reflects the high dimensionality of PFC activity19.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "The task subspace Q explained a smaller fraction of total variance in the PFC responses (11.0% and 7.0% for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC42 and reflects the high dimensionality of PFC activity19.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The task subspace Q explained a smaller fraction of total variance in the PFC responses (11.0% and 7.0% for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC42 and reflects the high dimensionality of PFC activity19.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC activity",
+      "label": "Measurement",
+      "sentence": "The task subspace Q explained a smaller fraction of total variance in the PFC responses (11.0% and 7.0% for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC42 and reflects the high dimensionality of PFC activity19.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "stimulus representations",
+      "label": "NeuralRepresentation",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "Mann–Whitney U-test",
+      "label": "StatisticalMethod",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "motion and color axes",
+      "label": "NeuralRepresentation",
+      "sentence": "This suppression was not due to correlation of activity along the motion and color axes with choice (Extended Data Figs. 7b and 8b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC activity",
+      "label": "Measurement",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "latent circuit connectivity",
+      "label": "Measurement",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "We confirmed that PFC responses significantly constrained the inferred latent connectivity above the effect of the task (Extended Data Figs. 7c and 8c and Methods), indicating that the suppression mechanism reflects dynamics in the PFC data.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC data",
+      "label": "DataType",
+      "sentence": "We confirmed that PFC responses significantly constrained the inferred latent connectivity above the effect of the task (Extended Data Figs. 7c and 8c and Methods), indicating that the suppression mechanism reflects dynamics in the PFC data.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC activity",
+      "label": "Measurement",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "regression methods",
+      "label": "StatisticalMethod",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC data",
+      "label": "DataType",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC data",
+      "label": "DataType",
+      "sentence": "To confirm that this difference results from recurrent interactions among task variables in the latent circuit model, we fitted PFC data with a modified latent circuit model in which the latent recurrent connectivity was constrained to be 0.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To confirm that this difference results from recurrent interactions among task variables in the latent circuit model, we fitted PFC data with a modified latent circuit model in which the latent recurrent connectivity was constrained to be 0.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "latent recurrent connectivity",
+      "label": "Measurement",
+      "sentence": "To confirm that this difference results from recurrent interactions among task variables in the latent circuit model, we fitted PFC data with a modified latent circuit model in which the latent recurrent connectivity was constrained to be 0.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "error trials",
+      "label": "BehavioralAssay",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "incongruent trials",
+      "label": "BehavioralAssay",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "relevant and irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "attention lapse",
+      "label": "Phenotype",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "To further test for behavioral relevance of representations identified by different dimensionality reduction methods, we again turned to RNNs in which we can measure behavioral effects of arbitrary activity perturbations.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "To further test for behavioral relevance of representations identified by different dimensionality reduction methods, we again turned to RNNs in which we can measure behavioral effects of arbitrary activity perturbations.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "StatisticalMethod",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "StatisticalMethod",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "decoder axis",
+      "label": "NeuralRepresentation",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "motion stimulus",
+      "label": "Stimulus",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Thus, as in the PFC, irrelevant sensory representations in our RNN appear not suppressed along the decoder axis, whereas they appear suppressed along the axis obtained from the latent circuit model (Fig. 7b).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "NeuralRepresentation",
+      "sentence": "Thus, as in the PFC, irrelevant sensory representations in our RNN appear not suppressed along the decoder axis, whereas they appear suppressed along the axis obtained from the latent circuit model (Fig. 7b).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Thus, as in the PFC, irrelevant sensory representations in our RNN appear not suppressed along the decoder axis, whereas they appear suppressed along the axis obtained from the latent circuit model (Fig. 7b).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "Thus, as in the PFC, irrelevant sensory representations in our RNN appear not suppressed along the decoder axis, whereas they appear suppressed along the axis obtained from the latent circuit model (Fig. 7b).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalModelComponent",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "StatisticalMethod",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "right motion stimulus",
+      "label": "Stimulus",
+      "sentence": "Specifically, stimulating the representation of the right motion stimulus should increase the proportion of right choices, shifting the decision boundary on motion context trials.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "Single neurons",
+      "label": "CellType",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "cognitive functions",
+      "label": "Phenomenon",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "single-neuron heterogeneity",
+      "label": "Measurement",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "dimensionality reduction",
+      "label": "Method",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Our theory for interpreting the latent circuit connectivity as a low-rank connectivity in RNNs enables causally validating low-dimensional circuit mechanisms via activity and connectivity perturbations in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "The latent circuit inference can be broadly applied to identify circuit mechanisms for different cognitive tasks from neural response data (Supplementary Fig. 7) and opens new possibilities for causally testing these mechanisms in future experiments.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "The latent circuit inference can be broadly applied to identify circuit mechanisms for different cognitive tasks from neural response data (Supplementary Fig. 7) and opens new possibilities for causally testing these mechanisms in future experiments.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural response data",
+      "label": "DataType",
+      "sentence": "The latent circuit inference can be broadly applied to identify circuit mechanisms for different cognitive tasks from neural response data (Supplementary Fig. 7) and opens new possibilities for causally testing these mechanisms in future experiments.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "latent dynamical systems",
+      "label": "ComputationalModel",
+      "sentence": "Although previous studies extensively modeled neural responses using various latent dynamical systems, our work demonstrates feasibility of fitting neural responses with low-dimensional recurrent circuits.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "Although previous studies extensively modeled neural responses using various latent dynamical systems, our work demonstrates feasibility of fitting neural responses with low-dimensional recurrent circuits.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "low-dimensional recurrent circuits",
+      "label": "ComputationalModel",
+      "sentence": "Although previous studies extensively modeled neural responses using various latent dynamical systems, our work demonstrates feasibility of fitting neural responses with low-dimensional recurrent circuits.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural data",
+      "label": "DataType",
+      "sentence": "Most methods fitting low-dimensional dynamical systems to neural data do not explicitly model task inputs and behavior and do not ground the inferred dynamics in the underlying network connectivity32,33,40,52,53.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "high-dimensional RNN architectures",
+      "label": "ComputationalModel",
+      "sentence": "On the other hand, high-dimensional RNN architectures serve as intermediate mechanistic models of neural responses, which generate latent dynamics from high-dimensional recurrent connectivity25–29.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "On the other hand, high-dimensional RNN architectures serve as intermediate mechanistic models of neural responses, which generate latent dynamics from high-dimensional recurrent connectivity25–29.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "The causal predictive power of the latent circuit model is further supported by its ability to predict RNN dynamics and behavior for out-of-distribution inputs (Supplementary Fig. 8).",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "RNN dynamics",
+      "label": "Phenomenon",
+      "sentence": "The causal predictive power of the latent circuit model is further supported by its ability to predict RNN dynamics and behavior for out-of-distribution inputs (Supplementary Fig. 8).",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The inhibitory suppression mechanism revealed by the latent circuit model is qualitatively similar to previous neural circuit models of how PFC flexibly switches between alternative stimulus–response mappings8–10,44,45.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "The inhibitory suppression mechanism revealed by the latent circuit model is qualitatively similar to previous neural circuit models of how PFC flexibly switches between alternative stimulus–response mappings8–10,44,45.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "stimulus–response mappings",
+      "label": "Phenomenon",
+      "sentence": "The inhibitory suppression mechanism revealed by the latent circuit model is qualitatively similar to previous neural circuit models of how PFC flexibly switches between alternative stimulus–response mappings8–10,44,45.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "Thus, RNNs do not find qualitatively distinct solutions to this task, and complex selectivity of single neurons has a simple explanation as a linear mixing of the low-dimensional latent inhibitory circuit mechanism.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "NeuralRepresentation",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "correlation-based dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "By contrast, correlation-based dimensionality reduction methods found no significant suppression of irrelevant stimuli in the same data19,21.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "By contrast, correlation-based dimensionality reduction methods found no significant suppression of irrelevant stimuli in the same data19,21.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "stimulus representations",
+      "label": "NeuralRepresentation",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "cognitive flexibility",
+      "label": "Phenomenon",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit model opens a route for interpreting circuit mechanisms in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "high-dimensional networks",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit model opens a route for interpreting circuit mechanisms in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "We fit the latent circuit model Eqs. (1)–(3) to neural response data y by minimizing the mean squared error loss function, L = ∑ k ∑ t ∥ ytk − Qxtk∥2+ ∥ ztk − woutxtk∥2, (7) using custom Python code59.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We fit the latent circuit model Eqs. (1)–(3) to neural response data y by minimizing the mean squared error loss function, L = ∑ k ∑ t ∥ ytk − Qxtk∥2+ ∥ ztk − woutxtk∥2, (7) using custom Python code59.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "neural response data",
+      "label": "DataType",
+      "sentence": "We fit the latent circuit model Eqs. (1)–(3) to neural response data y by minimizing the mean squared error loss function, L = ∑ k ∑ t ∥ ytk − Qxtk∥2+ ∥ ztk − woutxtk∥2, (7) using custom Python code59.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "mean squared error loss function",
+      "label": "StatisticalMethod",
+      "sentence": "We fit the latent circuit model Eqs. (1)–(3) to neural response data y by minimizing the mean squared error loss function, L = ∑ k ∑ t ∥ ytk − Qxtk∥2+ ∥ ztk − woutxtk∥2, (7) using custom Python code59.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "We perform this minimization using PyTorch and the Adam optimizer with default values 0.9 and 0.999 for the decay rate of the first and second moment estimates, respectively, a learning rate of 0.02 and a weight decay of 0.001.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "Adam optimizer",
+      "label": "Software",
+      "sentence": "We perform this minimization using PyTorch and the Adam optimizer with default values 0.9 and 0.999 for the decay rate of the first and second moment estimates, respectively, a learning rate of 0.02 and a weight decay of 0.001.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "Seaborn",
+      "label": "Software",
+      "sentence": "We used the Python software package Seaborn for visualizing model parameters and responses after training.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "We used the Python software package Seaborn for visualizing model parameters and responses after training.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "permutation test",
+      "label": "StatisticalMethod",
+      "sentence": "To determine whether the inferred latent circuit connectivity significantly depends on neural response data beyond the constraints imposed by the task alone, we performed a permutation test (Extended Data Figs. 1, 7c and 8c), which proceeds in three steps.",
+      "paper_location": "Methods: Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "latent circuit connectivity",
+      "label": "Measurement",
+      "sentence": "To determine whether the inferred latent circuit connectivity significantly depends on neural response data beyond the constraints imposed by the task alone, we performed a permutation test (Extended Data Figs. 1, 7c and 8c), which proceeds in three steps.",
+      "paper_location": "Methods: Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "neural response data",
+      "label": "DataType",
+      "sentence": "To determine whether the inferred latent circuit connectivity significantly depends on neural response data beyond the constraints imposed by the task alone, we performed a permutation test (Extended Data Figs. 1, 7c and 8c), which proceeds in three steps.",
+      "paper_location": "Methods: Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "Mann–Whitney U-test",
+      "label": "StatisticalMethod",
+      "sentence": "Third, we use a Mann–Whitney U-test to determine whether the correlation coefficients are significantly smaller for models fitted to the shuffled responses than original neural responses.",
+      "paper_location": "Methods: Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "Third, we use a Mann–Whitney U-test to determine whether the correlation coefficients are significantly smaller for models fitted to the shuffled responses than original neural responses.",
+      "paper_location": "Methods: Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "We used the same test for both RNN (N = 500; Extended Data Fig. 1) and PFC data (N = 1,000; Extended Data Figs. 7c and 8c).",
+      "paper_location": "Methods: Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "PFC data",
+      "label": "DataType",
+      "sentence": "We used the same test for both RNN (N = 500; Extended Data Fig. 1) and PFC data (N = 1,000; Extended Data Figs. 7c and 8c).",
+      "paper_location": "Methods: Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We consider RNNs of the form τẏ= −y + [Wrec y + Winu]+.",
+      "paper_location": "Methods: Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "rectified linear",
+      "label": "ActivationFunction",
+      "sentence": "Here, [⋅]+ is a rectified linear (ReLU) activation function, τ is a time constant, and u are external task inputs.",
+      "paper_location": "Methods: Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "ReLU",
+      "label": "ActivationFunction",
+      "sentence": "Here, [⋅]+ is a rectified linear (ReLU) activation function, τ is a time constant, and u are external task inputs.",
+      "paper_location": "Methods: Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "RNN simulation",
+      "label": "Method",
+      "sentence": "The RNN simulation and training were implemented in Python using the software package PyTorch.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "The RNN simulation and training were implemented in Python using the software package PyTorch.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "The RNN simulation and training were implemented in Python using the software package PyTorch.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "context cue",
+      "label": "Stimulus",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "sensory stimulus",
+      "label": "Stimulus",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion",
+      "label": "Stimulus",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "color features",
+      "label": "Stimulus",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "color coherence",
+      "label": "Measurement",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "stimuli",
+      "label": "Stimulus",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "Adam algorithm",
+      "label": "Software",
+      "sentence": "The training is performed with the Adam algorithm.",
+      "paper_location": "Methods: RNN training"
+    },
+    {
+      "entity": "linear regression model",
+      "label": "StatisticalMethod",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model c = βy + b, (27) where β ∈ ℝ1×N is the vector of regression coefficients, c ∈ ℝ1×K⋅T is the motion coherence on each trial, b ∈ ℝ is a bias term, and y ∈ ℝN×K⋅T is the RNN responses at each time step during the stimulus epoch of each trial.",
+      "paper_location": "Methods: Linear decoding"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model c = βy + b, (27) where β ∈ ℝ1×N is the vector of regression coefficients, c ∈ ℝ1×K⋅T is the motion coherence on each trial, b ∈ ℝ is a bias term, and y ∈ ℝN×K⋅T is the RNN responses at each time step during the stimulus epoch of each trial.",
+      "paper_location": "Methods: Linear decoding"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model c = βy + b, (27) where β ∈ ℝ1×N is the vector of regression coefficients, c ∈ ℝ1×K⋅T is the motion coherence on each trial, b ∈ ℝ is a bias term, and y ∈ ℝN×K⋅T is the RNN responses at each time step during the stimulus epoch of each trial.",
+      "paper_location": "Methods: Linear decoding"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "frontal eye field",
+      "label": "BrainRegion",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "neural activity recordings",
+      "label": "DataType",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "multiunits",
+      "label": "CellType",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "neural data",
+      "label": "DataType",
+      "sentence": "To facilitate comparison with previous studies analyzing the same dataset19,21, we used identical initial preprocessing of the neural data (using the publicly available code at https://www.ini.uzh.ch/en/research/groups/mante/data.html).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "saccade",
+      "label": "BehavioralResponse",
+      "sentence": "Monkeys reported their choice with a saccade to one of two targets presented shortly after fixation for the entire trial duration.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "fixation",
+      "label": "BehavioralAssay",
+      "sentence": "Monkeys reported their choice with a saccade to one of two targets presented shortly after fixation for the entire trial duration.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "random dots stimulus",
+      "label": "Stimulus",
+      "sentence": "We analyzed neural responses during the presentation of the random dots stimulus because the available data consisted of neural responses starting at 100 ms after stimulus onset for a duration of 750 ms.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "We analyzed neural responses during the presentation of the random dots stimulus because the available data consisted of neural responses starting at 100 ms after stimulus onset for a duration of 750 ms.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "stimulus onset",
+      "label": "StimulusTime",
+      "sentence": "We analyzed neural responses during the presentation of the random dots stimulus because the available data consisted of neural responses starting at 100 ms after stimulus onset for a duration of 750 ms.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "firing rates",
+      "label": "Measurement",
+      "sentence": "For each trial, we computed time-varying firing rates by counting spikes in a 50-ms sliding square window (50-ms steps).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "spikes",
+      "label": "Measurement",
+      "sentence": "For each trial, we computed time-varying firing rates by counting spikes in a 50-ms sliding square window (50-ms steps).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "Gaussian kernel",
+      "label": "StatisticalMethod",
+      "sentence": "Within the training and test sets, we z scored and smoothed (Gaussian kernel, σ = 40 ms) the response of each unit.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "Last, to denoise these trajectories, we projected them onto the principal components explaining 50% of their total variance (corresponding to the first 40 and 31 principal components for monkeys A and F, respectively).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "Last, to denoise these trajectories, we projected them onto the principal components explaining 50% of their total variance (corresponding to the first 40 and 31 principal components for monkeys A and F, respectively).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "PFC data",
+      "label": "DataType",
+      "sentence": "We fitted latent circuit models to the PFC data following similar procedures as for RNNs.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We fitted latent circuit models to the PFC data following similar procedures as for RNNs.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "National Institutes of Health",
+      "label": "FundingAgency",
+      "sentence": "This work was supported by National Institutes of Health grants RF1DA055666 and S10OD028632-01 (T.A.E.), the Alfred P. Sloan Foundation Research Fellowship (T.A.E.) and the Swartz Foundation Postdoctoral Fellowship (C.L.).",
+      "paper_location": "Acknowledgements"
+    },
+    {
+      "entity": "Alfred P. Sloan Foundation Research Fellowship",
+      "label": "FundingSource",
+      "sentence": "This work was supported by National Institutes of Health grants RF1DA055666 and S10OD028632-01 (T.A.E.), the Alfred P. Sloan Foundation Research Fellowship (T.A.E.) and the Swartz Foundation Postdoctoral Fellowship (C.L.).",
+      "paper_location": "Acknowledgements"
+    },
+    {
+      "entity": "Swartz Foundation Postdoctoral Fellowship",
+      "label": "FundingSource",
+      "sentence": "This work was supported by National Institutes of Health grants RF1DA055666 and S10OD028632-01 (T.A.E.), the Alfred P. Sloan Foundation Research Fellowship (T.A.E.) and the Swartz Foundation Postdoctoral Fellowship (C.L.).",
+      "paper_location": "Acknowledgements"
+    },
+    {
+      "entity": "MAP data-acquisition system",
+      "label": "Method",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "Plexon Inc.",
+      "label": "Organization",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "Spike sorting",
+      "label": "Method",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "principal component analysis",
+      "label": "Method",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "Plexon offline sorter",
+      "label": "Software",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "Plots were generated with Python (version 3.8) and Seaborn (version 0.9.0).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "Seaborn",
+      "label": "Software",
+      "sentence": "Plots were generated with Python (version 3.8) and Seaborn (version 0.9.0).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "Recurrent neural networks were trained using PyTorch (version 2.4).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "Recurrent neural networks",
+      "label": "ComputationalModel",
+      "sentence": "Recurrent neural networks were trained using PyTorch (version 2.4).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "GitHub",
+      "label": "SoftwareRepository",
+      "sentence": "The source code to reproduce results of this study is available on GitHub (https://github.com/engellab/latentcircuit) and Zenodo (https://zenodo.org/records/14020108).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "Zenodo",
+      "label": "SoftwareRepository",
+      "sentence": "The source code to reproduce results of this study is available on GitHub (https://github.com/engellab/latentcircuit) and Zenodo (https://zenodo.org/records/14020108).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "macaque monkeys",
+      "label": "Species",
+      "sentence": "Findings were successfully replicated in two macaque monkeys.",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "adult males",
+      "label": "SpeciesGroup",
+      "sentence": "The animals were adult males, ages 13 (monkey A) and 8 (monkey F) years old at the time the experiments were conducted.",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "National Institutes of Health",
+      "label": "FundingAgency",
+      "sentence": "All surgical and behavioral procedures conformed to the guidelines established by the National Institutes of Health and were approved by the Institutional Animal Care and Use Committee of Stanford University.",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "Institutional Animal Care and Use Committee of Stanford University",
+      "label": "EthicsCommittee",
+      "sentence": "All surgical and behavioral procedures conformed to the guidelines established by the National Institutes of Health and were approved by the Institutional Animal Care and Use Committee of Stanford University.",
+      "paper_location": "Reporting Summary"
+    }
+  ],
+  "key_terms": [
+    {
+      "term": "heterogeneous responses",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "task variables",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "correlations between neural activity and task variables",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "low-dimensional recurrent connectivity",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "suppression mechanism",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "patterned connectivity perturbations",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "causal interactions among task variables",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "behaviorally relevant computations",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "excitation and inhibition",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "circuit mechanism",
+      "sentence": "Thus, these models explicitly specify a circuit mechanism in the connectivity structure that gives rise to a dynamical mechanism controlling the flow of neural trajectories to implement task computations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "dynamical mechanism",
+      "sentence": "Thus, these models explicitly specify a circuit mechanism in the connectivity structure that gives rise to a dynamical mechanism controlling the flow of neural trajectories to implement task computations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "neural trajectories",
+      "sentence": "Thus, these models explicitly specify a circuit mechanism in the connectivity structure that gives rise to a dynamical mechanism controlling the flow of neural trajectories to implement task computations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "fixed points and linearized dynamics",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "low-dimensional latent variables",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "correlation-based methods",
+      "sentence": "However, unlike RNNs and circuit models, these correlation-based methods do not incorporate recurrent interactions among task variables, which implement computations necessary to solve the task.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "flexible trial-by-trial switching",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus–response mappings.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "low-dimensional circuit mechanisms",
+      "sentence": "However, because correlation-based dimensionality reduction methods bear no links to the connectivity and causal mechanisms, whether heterogeneous neural responses during cognitive tasks arise from low-dimensional circuit mechanisms remains an open question.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "low-dimensional latent circuit connectivity",
+      "sentence": "The model simultaneously infers a low-dimensional latent circuit connectivity generating task-relevant dynamics and heterogeneous mixing of these dynamics in single-neuron responses.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "task-relevant dynamics",
+      "sentence": "The model simultaneously infers a low-dimensional latent circuit connectivity generating task-relevant dynamics and heterogeneous mixing of these dynamics in single-neuron responses.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "heterogeneous mixing",
+      "sentence": "The model simultaneously infers a low-dimensional latent circuit connectivity generating task-relevant dynamics and heterogeneous mixing of these dynamics in single-neuron responses.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "task-related neural activity",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. (2).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "term": "low-dimensional subspace",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. (2).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "term": "orthonormal embedding matrix",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as y = Qx, (1) where Q ∈ ℝN×n is an orthonormal embedding matrix and n ≪ N.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "term": "connectivity matrices",
+      "sentence": "We can then derive an explicit relationship between connectivity matrices of a high-dimensional RNN and a low-dimensional latent circuit.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "term": "low-rank structure",
+      "sentence": "The relation Eq. (5) shows that the latent circuit connectivity wrec is a low-rank structure in the connectivity of a high-dimensional network, which captures interactions among the latent variables defined by the columns of Q.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "term": "invariant subspace",
+      "sentence": "Rather, it is a weaker condition that the linear subspace defined by Q is an invariant subspace of the high-dimensional recurrent connectivity matrix.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "term": "rank-one perturbation",
+      "sentence": "Specifically, a change in the connection δij between nodes i and j in the latent circuit maps onto a rank-one perturbation of the RNN connectivity matrix (Fig. 1e and Methods), δij → qiqT j , (6) where qi is the ith column of Q.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "term": "color context",
+      "sentence": "In the color context, the choice should be made according to the color and ignoring the motion stimulus and vice versa in the motion context.",
+      "paper_location": "Figure 2 legend"
+    },
+    {
+      "term": "motion context",
+      "sentence": "In the color context, the choice should be made according to the color and ignoring the motion stimulus and vice versa in the motion context.",
+      "paper_location": "Figure 2 legend"
+    },
+    {
+      "term": "psychometric functions",
+      "sentence": "Psychometric functions show that the RNN successfully learns the task; it responds to relevant stimuli and ignores irrelevant stimuli in each context.",
+      "paper_location": "Figure 2 legend"
+    },
+    {
+      "term": "decision boundary",
+      "sentence": "Indeed, weakening these connections in the latent circuit produced the predicted behavioral effect in the psychometric function, visible as a rotation of the decision boundary on motion context trials (Fig. 4a).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "term": "space of latent circuit solutions",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices wrec and projected the data onto the first two principal components, which accounted for 42% of total variance (Fig. 5a).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "term": "solution clusters",
+      "sentence": "The latent circuit solutions from all RNNs formed three major clusters.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "term": "asymmetric inhibitory mechanism",
+      "sentence": "RNNs in cluster 3 operate via a similar asymmetric inhibitory mechanism with the flipped biases for the left and right stimulus representations.",
+      "paper_location": "Extended Data Fig. 3 legend"
+    },
+    {
+      "term": "dynamical selection vector mechanism",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses19.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "term": "task subspace",
+      "sentence": "The task subspace Q explained a smaller fraction of total variance in the PFC responses (11.0% and 7.0% for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC42 and reflects the high dimensionality of PFC activity19.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "term": "task-relevant variance",
+      "sentence": "The task subspace Q explained a smaller fraction of total variance in the PFC responses (11.0% and 7.0% for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC42 and reflects the high dimensionality of PFC activity19.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "term": "error trials",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "term": "incongruent trials",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "term": "behavioral relevance",
+      "sentence": "To further test for behavioral relevance of representations identified by different dimensionality reduction methods, we again turned to RNNs in which we can measure behavioral effects of arbitrary activity perturbations.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "term": "decoder axis",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "term": "out-of-distribution inputs",
+      "sentence": "The causal predictive power of the latent circuit model is further supported by its ability to predict RNN dynamics and behavior for out-of-distribution inputs (Supplementary Fig. 8).",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "low-rank connectivity",
+      "sentence": "Our work explicitly relates low-dimensional recurrent circuits and low-rank connectivity in RNNs and enables inferring latent circuit structure in generic RNNs trained without low-rank connectivity constraints.",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "cognitive flexibility",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "Cayley transform",
+      "sentence": "To transform it into an unconstrained problem, we use the Cayley transform to parameterize orthonormal matrices by the linear space of skew symmetric matrices61, Q = (I + A)(I − A) −1 πn, (8) where πn represents projection onto the first n columns, and A is skew symmetric.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "term": "skew symmetric matrices",
+      "sentence": "To transform it into an unconstrained problem, we use the Cayley transform to parameterize orthonormal matrices by the linear space of skew symmetric matrices61, Q = (I + A)(I − A) −1 πn, (8) where πn represents projection onto the first n columns, and A is skew symmetric.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "term": "minibatch size",
+      "sentence": "We use a minibatch size of 128 trials.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "term": "random initializations",
+      "sentence": "For each of these RNNs, we fitted an ensemble of 100 latent circuit models starting with random initializations of the parameters for latent connectivity and embedding Q and selected 10 latent circuits with the highest fit quality on the test data, which formed the set of converged solutions (Methods).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "term": "Dale’s law",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale’s law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "term": "stimulus coherence",
+      "sentence": "The strength of sensory evidence for motion and color varies randomly from trial to trial controlled by the stimulus coherence.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "term": "trial-averaged neural responses",
+      "sentence": "We fitted the latent circuit model to trial-averaged neural responses on correct trials.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "term": "z scored and smoothed",
+      "sentence": "Within the training and test sets, we z scored and smoothed (Gaussian kernel, σ = 40 ms) the response of each unit.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "term": "Nature Portfolio Reporting Summary",
+      "sentence": "Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.",
+      "paper_location": "Methods: Reporting summary"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_gpt-5.5_20260629_090841.json b/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_gpt-5.5_20260629_090841.json
new file mode 100644
index 0000000..883572e
--- /dev/null
+++ b/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_gpt-5.5_20260629_090841.json
@@ -0,0 +1,1804 @@
+{
+  "source_metadata": {
+    "source_path": "evaluation/ner/latent-circuit-inference/s41593-025-01869-7.pdf",
+    "statistics": {
+      "model": "gpt-5.5",
+      "prompt_file": "/Users/pujatrivedi/Desktop/MIT/structsense/evaluation/ner/direct_llm_call/prompts/extractor_neuroscience_ner.txt",
+      "extracted_at": "20260629_090841",
+      "total_entities": 251,
+      "entities_by_label": {
+        "Method": 30,
+        "ComputationalModel": 26,
+        "Measurement": 25,
+        "Phenomenon": 22,
+        "BehavioralAssay": 20,
+        "BrainRegion": 17,
+        "ComputationalVariable": 16,
+        "TaskVariable": 14,
+        "NeuralResponse": 14,
+        "Species": 12,
+        "CellType": 10,
+        "Stimulus": 9,
+        "Software": 8,
+        "Dataset": 7,
+        "NeuralConnectivity": 6,
+        "NeuralRepresentation": 6,
+        "ComputationalUnit": 3,
+        "NeuralSignal": 2,
+        "Organization": 2,
+        "NeuralCircuit": 1,
+        "Instrument": 1
+      }
+    }
+  },
+  "entities": [
+    {
+      "entity": "Latent circuit inference",
+      "label": "Method",
+      "sentence": "Latent circuit inference from heterogeneous neural responses during cognitive tasks",
+      "paper_location": "Title"
+    },
+    {
+      "entity": "heterogeneous neural responses",
+      "label": "Phenomenon",
+      "sentence": "Latent circuit inference from heterogeneous neural responses during cognitive tasks",
+      "paper_location": "Title"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Latent circuit inference from heterogeneous neural responses during cognitive tasks",
+      "paper_location": "Title"
+    },
+    {
+      "entity": "Higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "sensory, cognitive and motor signals",
+      "label": "NeuralSignal",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "heterogeneous responses",
+      "label": "Phenomenon",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "task variables",
+      "label": "TaskVariable",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "Dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "task variables",
+      "label": "TaskVariable",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "heterogeneous responses",
+      "label": "Phenomenon",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "dimensionality reduction approach",
+      "label": "Method",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "task variables",
+      "label": "TaskVariable",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "low-dimensional recurrent connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "recurrent neural networks",
+      "label": "ComputationalModel",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "contextual representations",
+      "label": "NeuralRepresentation",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "NeuralResponse",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "patterned connectivity perturbations",
+      "label": "Method",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "NeuralResponse",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "task variables",
+      "label": "TaskVariable",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Dataset",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "Cognitive functions",
+      "label": "Phenomenon",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "sensory and contextual signals",
+      "label": "NeuralSignal",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral response",
+      "label": "Measurement",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "excitatory and inhibitory neurons",
+      "label": "CellType",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cortical circuits",
+      "label": "NeuralCircuit",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "excitation and inhibition",
+      "label": "Phenomenon",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural populations",
+      "label": "CellType",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "task variables",
+      "label": "TaskVariable",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "dynamical system",
+      "label": "ComputationalModel",
+      "sentence": "Because these circuit models usually assume a relatively simple connectivity structure, their connectivity can be directly related to a dynamical system description of computations supporting cognitive task execution5,6.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cognitive task execution",
+      "label": "BehavioralAssay",
+      "sentence": "Because these circuit models usually assume a relatively simple connectivity structure, their connectivity can be directly related to a dynamical system description of computations supporting cognitive task execution5,6.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural dynamics",
+      "label": "Phenomenon",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation–inhibition balance12) and thus can be causally validated in experiments13–15.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation–inhibition balance12) and thus can be causally validated in experiments13–15.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral performance",
+      "label": "Measurement",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation–inhibition balance12) and thus can be causally validated in experiments13–15.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "excitation–inhibition balance",
+      "label": "Phenomenon",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation–inhibition balance12) and thus can be causally validated in experiments13–15.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "NeuralResponse",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of single-neuron responses in the cortex.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cortex",
+      "label": "BrainRegion",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of single-neuron responses in the cortex.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "Single neurons",
+      "label": "CellType",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "heterogeneous tuning",
+      "label": "Phenomenon",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "task variables",
+      "label": "TaskVariable",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "recurrent neural network",
+      "label": "ComputationalModel",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural activity data",
+      "label": "Dataset",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks30.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks30.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "fixed points",
+      "label": "Phenomenon",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "linearized dynamics",
+      "label": "Phenomenon",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "dynamical system",
+      "label": "ComputationalModel",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Dataset",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural population activity",
+      "label": "Measurement",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "low-dimensional latent variables",
+      "label": "ComputationalVariable",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "unsupervised methods",
+      "label": "Method",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution26,37–40.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "task inputs",
+      "label": "TaskVariable",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution26,37–40.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral outputs",
+      "label": "Measurement",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution26,37–40.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "targeted dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural representations",
+      "label": "NeuralRepresentation",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "task inputs",
+      "label": "TaskVariable",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral outputs",
+      "label": "Measurement",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural population activity",
+      "label": "Measurement",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC’s role in context-dependent decision-making.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC’s role in context-dependent decision-making.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "Context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus–response mappings.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "stimulus–response mappings",
+      "label": "TaskVariable",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus–response mappings.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "NeuralRepresentation",
+      "sentence": "Most circuit models hypothesize a relatively simple mechanism based on inhibition of sensory representations irrelevant in the current context8–10,44–46.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "NeuralResponse",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural responses",
+      "label": "NeuralResponse",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "task variables",
+      "label": "TaskVariable",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "low-dimensional latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "heterogeneous neural responses",
+      "label": "Phenomenon",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "irrelevant sensory representations",
+      "label": "NeuralRepresentation",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "single-neuron heterogeneity",
+      "label": "Phenomenon",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "high-dimensional neural responses",
+      "label": "NeuralResponse",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neurons",
+      "label": "CellType",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "cognitive task",
+      "label": "BehavioralAssay",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "low-dimensional latent variables",
+      "label": "ComputationalVariable",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "ReLU",
+      "label": "Method",
+      "sentence": "where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "rectified linear",
+      "label": "Method",
+      "sentence": "where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "task inputs",
+      "label": "TaskVariable",
+      "sentence": "The latent nodes interact via the recurrent connectivity wrec and receive external task inputs u through the input connectivity win.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "task outputs",
+      "label": "TaskVariable",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity wout:",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "circuit activity",
+      "label": "Measurement",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity wout:",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "NeuralResponse",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "task inputs",
+      "label": "TaskVariable",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "cognitive task",
+      "label": "BehavioralAssay",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "mixed selectivity",
+      "label": "Phenomenon",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "neural responses",
+      "label": "NeuralResponse",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "cortex",
+      "label": "BrainRegion",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We consider RNNs with dynamics",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "The relation between connectivity matrices has the powerful consequence that we can validate the latent circuit mechanism directly in the RNN connectivity.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "The relation between connectivity matrices has the powerful consequence that we can validate the latent circuit mechanism directly in the RNN connectivity.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit mechanism",
+      "label": "ComputationalModel",
+      "sentence": "The validation of inferred circuit mechanisms via RNN perturbations is critical because the fit quality alone does not guarantee that the inferred model captures the correct mechanism that generated data40.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "Latent circuit for context-dependent decision-making",
+      "paper_location": "Results heading"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory stimulus",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context cue",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalUnit",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "heterogeneous mixed selectivity",
+      "label": "Phenomenon",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "task variables",
+      "label": "TaskVariable",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalVariable",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory color nodes",
+      "label": "ComputationalVariable",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory motion nodes",
+      "label": "ComputationalVariable",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "choice nodes",
+      "label": "ComputationalVariable",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "This choice of the latent circuit dimensionality agrees with the dimensionality of RNN responses after training, which is usually close to the total number of inputs and outputs (the first eight principal components accounted for 97.9% of the total variance in RNN responses; Supplementary Fig. 2).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "NeuralResponse",
+      "sentence": "This choice of the latent circuit dimensionality agrees with the dimensionality of RNN responses after training, which is usually close to the total number of inputs and outputs (the first eight principal components accounted for 97.9% of the total variance in RNN responses; Supplementary Fig. 2).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalVariable",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "stimuli",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "excitatory connections",
+      "label": "NeuralConnectivity",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context node",
+      "label": "ComputationalVariable",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "inhibitory connections",
+      "label": "NeuralConnectivity",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalVariable",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion context node",
+      "label": "ComputationalVariable",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "This pattern of connections from the context nodes to the sensory nodes implements a suppression mechanism that inhibits the irrelevant stimulus–response mapping in each context.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant sensory representation",
+      "label": "NeuralRepresentation",
+      "sentence": "Because the irrelevant sensory representation is suppressed, it does not drive the decision output.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "decision output",
+      "label": "Measurement",
+      "sentence": "Because the irrelevant sensory representation is suppressed, it does not drive the decision output.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "First, we verified the signatures of the suppression mechanism in the RNN activity.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "First, we verified the signatures of the suppression mechanism in the RNN activity.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "NeuralResponse",
+      "sentence": "We projected RNN responses onto the columns of Q, which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity patterns",
+      "label": "Measurement",
+      "sentence": "We projected RNN responses onto the columns of Q, which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "choice axis",
+      "label": "ComputationalVariable",
+      "sentence": "Next, the choice axis is the difference of two columns of Q corresponding to the left and right choice nodes in the latent circuit.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion axis",
+      "label": "ComputationalVariable",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color axis",
+      "label": "ComputationalVariable",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "Projections of RNN activity onto the motion and color axes reveal representations of relevant sensory stimuli, whereas representations of irrelevant stimuli are suppressed.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralConnectivity",
+      "sentence": "Second, we used the connectivity relationships Eq. (5) to directly validate the latent circuit mechanism in the RNN connectivity.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "psychometric function",
+      "label": "Measurement",
+      "sentence": "Indeed, weakening these connections in the latent circuit produced the predicted behavioral effect in the psychometric function, visible as a rotation of the decision boundary on motion context trials (Fig. 4a).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We next asked whether RNNs trained on the same task arrive at different circuit solutions for context-dependent decision-making.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "We next asked whether RNNs trained on the same task arrive at different circuit solutions for context-dependent decision-making.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "principal component analysis",
+      "label": "Method",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices wrec and projected the data onto the first two principal components, which accounted for 42% of total variance (Fig. 5a).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices wrec and projected the data onto the first two principal components, which accounted for 42% of total variance (Fig. 5a).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "Gaussian mixture model",
+      "label": "Method",
+      "sentence": "The projection reveals three major solution clusters, which we partitioned by fitting data with a Gaussian mixture model with three components.",
+      "paper_location": "Figure 5 caption"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Representations of irrelevant stimuli in the PFC",
+      "paper_location": "Results heading"
+    },
+    {
+      "entity": "irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "Representations of irrelevant stimuli in the PFC",
+      "paper_location": "Results heading"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "NeuralResponse",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "regression models",
+      "label": "Method",
+      "sentence": "This conclusion was derived using dimensionality reduction methods that fit neural responses with regression models19,21,41–43 to find low-dimensional projections that best correlate with task variables (Fig. 1b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC neurons",
+      "label": "CellType",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkeys A and F",
+      "label": "Species",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "rhesus monkeys",
+      "label": "Species",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "smoothed condition-averaged PFC responses",
+      "label": "NeuralResponse",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting 100 ms after the stimulus onset19 (Methods).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "stimulus onset",
+      "label": "Stimulus",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting 100 ms after the stimulus onset19 (Methods).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "NeuralResponse",
+      "sentence": "The latent circuit model provided good fits of PFC responses projected onto the low-dimensional subspace spanned by the columns of the inferred embedding matrix Q for both monkeys (r2 = 0.88 and 0.76 on test data for monkeys A and F, respectively; Extended Data Figs. 7a and 8a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkeys A and F",
+      "label": "Species",
+      "sentence": "The latent circuit model provided good fits of PFC responses projected onto the low-dimensional subspace spanned by the columns of the inferred embedding matrix Q for both monkeys (r2 = 0.88 and 0.76 on test data for monkeys A and F, respectively; Extended Data Figs. 7a and 8a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The task subspace Q explained a smaller fraction of total variance in the PFC responses (11.0% and 7.0% for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC42 and reflects the high dimensionality of PFC activity19.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC activity",
+      "label": "Measurement",
+      "sentence": "The task subspace Q explained a smaller fraction of total variance in the PFC responses (11.0% and 7.0% for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC42 and reflects the high dimensionality of PFC activity19.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "Mann–Whitney U-test",
+      "label": "Method",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "stimulus representations",
+      "label": "NeuralRepresentation",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "NeuralResponse",
+      "sentence": "We confirmed that PFC responses significantly constrained the inferred latent connectivity above the effect of the task (Extended Data Figs. 7c and 8c and Methods), indicating that the suppression mechanism reflects dynamics in the PFC data.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC data",
+      "label": "Dataset",
+      "sentence": "We confirmed that PFC responses significantly constrained the inferred latent connectivity above the effect of the task (Extended Data Figs. 7c and 8c and Methods), indicating that the suppression mechanism reflects dynamics in the PFC data.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "error trials",
+      "label": "BehavioralAssay",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "To further test for behavioral relevance of representations identified by different dimensionality reduction methods, we again turned to RNNs in which we can measure behavioral effects of arbitrary activity perturbations.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "activity perturbations",
+      "label": "Method",
+      "sentence": "To further test for behavioral relevance of representations identified by different dimensionality reduction methods, we again turned to RNNs in which we can measure behavioral effects of arbitrary activity perturbations.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "Method",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "decoder axis",
+      "label": "ComputationalVariable",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "irrelevant motion stimulus",
+      "label": "Stimulus",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "single-neuron heterogeneity",
+      "label": "Phenomenon",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "high-dimensional networks",
+      "label": "ComputationalModel",
+      "sentence": "We show that low-dimensional circuit mechanisms can explain task-relevant dynamics in high-dimensional networks and establish feasibility of inferring these mechanisms from neural response data.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Dataset",
+      "sentence": "We show that low-dimensional circuit mechanisms can explain task-relevant dynamics in high-dimensional networks and establish feasibility of inferring these mechanisms from neural response data.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "activity and connectivity perturbations",
+      "label": "Method",
+      "sentence": "Our theory for interpreting the latent circuit connectivity as a low-rank connectivity in RNNs enables causally validating low-dimensional circuit mechanisms via activity and connectivity perturbations in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural population dynamics",
+      "label": "Phenomenon",
+      "sentence": "Although previous studies extensively modeled neural responses using various latent dynamical systems, our work demonstrates feasibility of fitting neural responses with low-dimensional recurrent circuits.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "single-trial neural activity",
+      "label": "Measurement",
+      "sentence": "Our results motivate future work incorporating low-dimensional recurrent circuits as latent dynamics generators within these model architectures, which will enable combining their ability of fitting single-trial neural activity26,29 with the uniqueness and interpretability furnished by the latent circuit model.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural dynamics",
+      "label": "Phenomenon",
+      "sentence": "By relating connectivity to neural dynamics and behavior, the latent circuit model extends the causal predictive power of classical neural circuit models to the study of mechanisms of cognitive functions in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "By relating connectivity to neural dynamics and behavior, the latent circuit model extends the causal predictive power of classical neural circuit models to the study of mechanisms of cognitive functions in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "line attractor",
+      "label": "Phenomenon",
+      "sentence": "In this mechanism, choice is integrated along a line attractor in each context (l, choice axis).",
+      "paper_location": "Extended Data Fig. 4"
+    },
+    {
+      "entity": "choice axis",
+      "label": "ComputationalVariable",
+      "sentence": "In this mechanism, choice is integrated along a line attractor in each context (l, choice axis).",
+      "paper_location": "Extended Data Fig. 4"
+    },
+    {
+      "entity": "selection vector",
+      "label": "ComputationalVariable",
+      "sentence": "Only inputs aligned with the left eigenvector (vL, selection vector) drive activity along the right eigenvector (vR), which has a non-zero projection onto the choice axis l.",
+      "paper_location": "Extended Data Fig. 4"
+    },
+    {
+      "entity": "Softplus",
+      "label": "Method",
+      "sentence": "We trained RNNs with the Softplus activation function of units to perform the context-dependent decision-making task, using different values of the parameter β across RNNs (25 RNNs for each value of β).",
+      "paper_location": "Extended Data Fig. 5"
+    },
+    {
+      "entity": "ReLU",
+      "label": "Method",
+      "sentence": "The latent connectivity reveals a similar inhibitory mechanism as found in RNNs with the ReLU activation function.",
+      "paper_location": "Extended Data Fig. 5"
+    },
+    {
+      "entity": "PFC recordings",
+      "label": "Dataset",
+      "sentence": "Latent circuit analysis of PFC recordings from monkey A.",
+      "paper_location": "Extended Data Fig. 7"
+    },
+    {
+      "entity": "monkey A",
+      "label": "Species",
+      "sentence": "Latent circuit analysis of PFC recordings from monkey A.",
+      "paper_location": "Extended Data Fig. 7"
+    },
+    {
+      "entity": "PFC recordings",
+      "label": "Dataset",
+      "sentence": "Latent circuit analysis of PFC recordings from monkey F.",
+      "paper_location": "Extended Data Fig. 8"
+    },
+    {
+      "entity": "monkey F",
+      "label": "Species",
+      "sentence": "Latent circuit analysis of PFC recordings from monkey F.",
+      "paper_location": "Extended Data Fig. 8"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "We fit the latent circuit model Eqs. (1)–(3) to neural response data y by minimizing the mean squared error loss function, using custom Python code59.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "We perform this minimization using PyTorch and the Adam optimizer with default values 0.9 and 0.999 for the decay rate of the first and second moment estimates, respectively, a learning rate of 0.02 and a weight decay of 0.001.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "Adam optimizer",
+      "label": "Software",
+      "sentence": "We perform this minimization using PyTorch and the Adam optimizer with default values 0.9 and 0.999 for the decay rate of the first and second moment estimates, respectively, a learning rate of 0.02 and a weight decay of 0.001.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "Seaborn",
+      "label": "Software",
+      "sentence": "We used the Python software package Seaborn for visualizing model parameters and responses after training.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "Mann–Whitney U-test",
+      "label": "Method",
+      "sentence": "Third, we use a Mann–Whitney U-test to determine whether the correlation coefficients are significantly smaller for models fitted to the shuffled responses than original neural responses.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "ReLU",
+      "label": "Method",
+      "sentence": "Here, [⋅]+ is a rectified linear (ReLU) activation function, τ is a time constant, and u are external task inputs.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "Softplus",
+      "label": "Method",
+      "sentence": "To test whether our results extend to networks with a different biologically plausible nonlinearity, we trained RNNs that had a Softplus activation function f(x) = g β log(1 + eβx) for a range of parameter β and also with varying gain g across units.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "Euler scheme",
+      "label": "Method",
+      "sentence": "We discretize the RNN dynamics Eq. (10) using the first-order Euler scheme with a time step Δt and add a noise term to obtain",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "Dale’s law",
+      "label": "Phenomenon",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale’s law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "excitatory units",
+      "label": "ComputationalUnit",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale’s law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "inhibitory units",
+      "label": "ComputationalUnit",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale’s law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "Context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "Context-dependent decision-making task",
+      "paper_location": "Methods heading"
+    },
+    {
+      "entity": "context cue",
+      "label": "Stimulus",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "sensory stimulus",
+      "label": "Stimulus",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "color coherence",
+      "label": "Measurement",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "linear regression model",
+      "label": "Method",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "frontal eye field",
+      "label": "BrainRegion",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "multiunits",
+      "label": "CellType",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "spikes",
+      "label": "Measurement",
+      "sentence": "For each trial, we computed time-varying firing rates by counting spikes in a 50-ms sliding square window (50-ms steps).",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "firing rates",
+      "label": "Measurement",
+      "sentence": "For each trial, we computed time-varying firing rates by counting spikes in a 50-ms sliding square window (50-ms steps).",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "Gaussian kernel",
+      "label": "Method",
+      "sentence": "Within the training and test sets, we z scored and smoothed (Gaussian kernel, σ = 40 ms) the response of each unit.",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "Last, to denoise these trajectories, we projected them onto the principal components explaining 50% of their total variance (corresponding to the first 40 and 31 principal components for monkeys A and F, respectively).",
+      "paper_location": "Methods"
+    },
+    {
+      "entity": "National Institutes of Health",
+      "label": "Organization",
+      "sentence": "This work was supported by National Institutes of Health grants RF1DA055666 and S10OD028632-01 (T.A.E.), the Alfred P. Sloan Foundation Research Fellowship (T.A.E.) and the Swartz Foundation Postdoctoral Fellowship (C.L.).",
+      "paper_location": "Acknowledgements"
+    },
+    {
+      "entity": "macaque monkeys",
+      "label": "Species",
+      "sentence": "Findings were successfully replicated in two macaque monkeys.",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "MAP data-acquisition system",
+      "label": "Instrument",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "Plexon Inc.",
+      "label": "Organization",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "Spike sorting",
+      "label": "Method",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "principal component analysis",
+      "label": "Method",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "Plexon offline sorter",
+      "label": "Software",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "Recurrent neural networks were trained using PyTorch (version 2.4).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "GitHub",
+      "label": "Software",
+      "sentence": "The source code to reproduce results of this study is available on GitHub (https://github.com/engellab/latentcircuit) and Zenodo (https://zenodo.org/records/14020108).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "Zenodo",
+      "label": "Software",
+      "sentence": "The source code to reproduce results of this study is available on GitHub (https://github.com/engellab/latentcircuit) and Zenodo (https://zenodo.org/records/14020108).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "adult males",
+      "label": "Species",
+      "sentence": "The animals were adult males, ages 13 (monkey A) and 8 (monkey F) years old at the time the experiments were conducted.",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "monkey A",
+      "label": "Species",
+      "sentence": "The animals were adult males, ages 13 (monkey A) and 8 (monkey F) years old at the time the experiments were conducted.",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "entity": "monkey F",
+      "label": "Species",
+      "sentence": "The animals were adult males, ages 13 (monkey A) and 8 (monkey F) years old at the time the experiments were conducted.",
+      "paper_location": "Reporting Summary"
+    }
+  ],
+  "key_terms": [
+    {
+      "term": "latent circuit inference",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "context-dependent decision-making task",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "patterned connectivity perturbations",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "causal interactions among task variables",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "hand-crafted neural circuit models",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "excitation–inhibition balance",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation–inhibition balance12) and thus can be causally validated in experiments13–15.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "single-neuron responses",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of single-neuron responses in the cortex.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "mixed selectivity",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "term": "low-rank structure",
+      "sentence": "The relation Eq. (5) shows that the latent circuit connectivity wrec is a low-rank structure in the connectivity of a high-dimensional network, which captures interactions among the latent variables defined by the columns of Q.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "term": "rank-one perturbation",
+      "sentence": "Specifically, a change in the connection δij between nodes i and j in the latent circuit maps onto a rank-one perturbation of the RNN connectivity matrix (Fig. 1e and Methods),",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "term": "context cue",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results"
+    },
+    {
+      "term": "motion coherence",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods"
+    },
+    {
+      "term": "color coherence",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods"
+    },
+    {
+      "term": "alternative stimulus–response mappings",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus–response mappings.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "suppression mechanism",
+      "sentence": "This pattern of connections from the context nodes to the sensory nodes implements a suppression mechanism that inhibits the irrelevant stimulus–response mapping in each context.",
+      "paper_location": "Results"
+    },
+    {
+      "term": "inhibition of irrelevant sensory representations",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "low-dimensional circuit mechanism",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "correlation-based dimensionality reduction methods",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC’s role in context-dependent decision-making.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "targeted dimensionality reduction methods",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "population state space",
+      "sentence": "Each column of D defines an axis in the neural population state space such that the projection of neural activity onto this axis correlates with a specific task variable (right).",
+      "paper_location": "Figure 1 caption"
+    },
+    {
+      "term": "invariant subspace",
+      "sentence": "This equation states that the subspace spanned by the columns of Q is an invariant subspace of the high-dimensional system; that is, the vector field at any point in this subspace lies entirely in this subspace.",
+      "paper_location": "Figure 1 caption"
+    },
+    {
+      "term": "psychometric functions",
+      "sentence": "Psychometric functions show that the RNN successfully learns the task; it responds to relevant stimuli and ignores irrelevant stimuli in each context.",
+      "paper_location": "Figure 2 caption"
+    },
+    {
+      "term": "choice axis",
+      "sentence": "Next, the choice axis is the difference of two columns of Q corresponding to the left and right choice nodes in the latent circuit.",
+      "paper_location": "Results"
+    },
+    {
+      "term": "motion axis",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results"
+    },
+    {
+      "term": "color axis",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results"
+    },
+    {
+      "term": "motion context trials",
+      "sentence": "Indeed, weakening these connections in the latent circuit produced the predicted behavioral effect in the psychometric function, visible as a rotation of the decision boundary on motion context trials (Fig. 4a).",
+      "paper_location": "Results"
+    },
+    {
+      "term": "decision boundary",
+      "sentence": "Indeed, weakening these connections in the latent circuit produced the predicted behavioral effect in the psychometric function, visible as a rotation of the decision boundary on motion context trials (Fig. 4a).",
+      "paper_location": "Results"
+    },
+    {
+      "term": "space of circuit mechanisms",
+      "sentence": "To explore the space of circuit mechanisms, we trained an ensemble of 200 RNNs with randomly initialized connectivity to the same level of task performance (r2 = 0.93 ± 0.01, coefficient of determination for the RNN and target outputs, mean ± s.d. across networks; Supplementary Fig. 4).",
+      "paper_location": "Results"
+    },
+    {
+      "term": "converged solutions",
+      "sentence": "For each of these RNNs, we fitted an ensemble of 100 latent circuit models starting with random initializations of the parameters for latent connectivity and embedding Q and selected 10 latent circuits with the highest fit quality on the test data, which formed the set of converged solutions (Methods).",
+      "paper_location": "Results"
+    },
+    {
+      "term": "Gaussian mixture model with three components",
+      "sentence": "The projection reveals three major solution clusters, which we partitioned by fitting data with a Gaussian mixture model with three components.",
+      "paper_location": "Figure 5 caption"
+    },
+    {
+      "term": "asymmetric inhibitory mechanism",
+      "sentence": "Asymmetric inhibitory mechanism for context-dependent decision-making.",
+      "paper_location": "Extended Data Fig. 3"
+    },
+    {
+      "term": "dynamical selection vector mechanism",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses19.",
+      "paper_location": "Results"
+    },
+    {
+      "term": "error trials",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results"
+    },
+    {
+      "term": "incongruent trials",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results"
+    },
+    {
+      "term": "linear decoder",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results"
+    },
+    {
+      "term": "decoder axis",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results"
+    },
+    {
+      "term": "activity patterns",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results"
+    },
+    {
+      "term": "task-relevant dynamics",
+      "sentence": "We show that low-dimensional circuit mechanisms can explain task-relevant dynamics in high-dimensional networks and establish feasibility of inferring these mechanisms from neural response data.",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "latent dynamical systems",
+      "sentence": "Although previous studies extensively modeled neural responses using various latent dynamical systems, our work demonstrates feasibility of fitting neural responses with low-dimensional recurrent circuits.",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "high-dimensional RNN architectures",
+      "sentence": "On the other hand, high-dimensional RNN architectures serve as intermediate mechanistic models of neural responses, which generate latent dynamics from high-dimensional recurrent connectivity25–29.",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "low-rank connectivity",
+      "sentence": "Our work explicitly relates low-dimensional recurrent circuits and low-rank connectivity in RNNs and enables inferring latent circuit structure in generic RNNs trained without low-rank connectivity constraints.",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "cognitive flexibility",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "Cayley transform",
+      "sentence": "To transform it into an unconstrained problem, we use the Cayley transform to parameterize orthonormal matrices by the linear space of skew symmetric matrices61,",
+      "paper_location": "Methods"
+    },
+    {
+      "term": "nonlinear least squares optimization",
+      "sentence": "Because the variable x depends implicitly on the latent circuit parameters wrec and win, the minimization of L is a nonlinear least squares optimization problem60 in which we simultaneously search for a behaviorally relevant projection of the high-dimensional activity and a low-dimensional neural circuit that generates dynamics in this projection.",
+      "paper_location": "Methods"
+    },
+    {
+      "term": "permutation test",
+      "sentence": "To determine whether the inferred latent circuit connectivity significantly depends on neural response data beyond the constraints imposed by the task alone, we performed a permutation test (Extended Data Figs. 1, 7c and 8c), which proceeds in three steps.",
+      "paper_location": "Methods"
+    },
+    {
+      "term": "shuffling procedure",
+      "sentence": "Our shuffling procedure randomly permutes neural responses with respect to trial conditions while preserving the input–output relationship on each trial so that the fitted latent circuit models can still perform the task.",
+      "paper_location": "Methods"
+    },
+    {
+      "term": "Dale’s law",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale’s law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods"
+    },
+    {
+      "term": "first-order Euler scheme",
+      "sentence": "We discretize the RNN dynamics Eq. (10) using the first-order Euler scheme with a time step Δt and add a noise term to obtain",
+      "paper_location": "Methods"
+    },
+    {
+      "term": "mean squared error",
+      "sentence": "To train the RNN, we minimize the mean squared error between the output z(t) and the target ztarget(t):",
+      "paper_location": "Methods"
+    },
+    {
+      "term": "L2 penalty",
+      "sentence": "We control the degree of correlation between the input and output vectors in the RNN by adding an L2 penalty",
+      "paper_location": "Methods"
+    },
+    {
+      "term": "spike sorting",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary"
+    },
+    {
+      "term": "MAP data-acquisition system",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_gpt-5.5_20260629_092216.json b/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_gpt-5.5_20260629_092216.json
new file mode 100644
index 0000000..0f6a4b7
--- /dev/null
+++ b/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_gpt-5.5_20260629_092216.json
@@ -0,0 +1,2341 @@
+{
+  "source_metadata": {
+    "source_path": "evaluation/ner/latent-circuit-inference/s41593-025-01869-7.pdf",
+    "statistics": {
+      "model": "gpt-5.5",
+      "prompt_file": "/Users/pujatrivedi/Desktop/MIT/structsense/evaluation/ner/direct_llm_call/prompts/extractor_neuroscience_ner.txt",
+      "temperature": null,
+      "seed": null,
+      "response_status": "completed",
+      "extracted_at": "20260629_092216",
+      "total_entities": 336,
+      "entities_by_label": {
+        "Measurement": 56,
+        "Method": 35,
+        "ComputationalEntity": 35,
+        "ModelArchitecture": 27,
+        "Stimulus": 21,
+        "BrainRegion": 19,
+        "Species": 17,
+        "BehavioralAssay": 17,
+        "ExperimentalVariable": 15,
+        "DataType": 14,
+        "Software": 13,
+        "CellType": 11,
+        "Phenomenon": 7,
+        "NeuralRepresentation": 6,
+        "ComputationalModel": 6,
+        "SignalType": 5,
+        "ExperimentalCondition": 5,
+        "ComputationalMethod": 4,
+        "StimulusFeature": 4,
+        "Organization": 4,
+        "NeuralCircuit": 3,
+        "BehavioralResponse": 2,
+        "SoftwareRepository": 2,
+        "Grant": 2,
+        "CognitiveFunction": 1,
+        "ExperimentalParadigm": 1,
+        "StimulusTiming": 1,
+        "Phenotype": 1,
+        "Instrument": 1,
+        "OrganismAttribute": 1
+      }
+    }
+  },
+  "entities": [
+    {
+      "entity": "Higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "sensory",
+      "label": "SignalType",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "cognitive",
+      "label": "SignalType",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "motor signals",
+      "label": "SignalType",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "task variables",
+      "label": "ExperimentalVariable",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "Dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "task variables",
+      "label": "ExperimentalVariable",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "connectivity",
+      "label": "Measurement",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "Method",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "task variables",
+      "label": "ExperimentalVariable",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "low-dimensional recurrent connectivity",
+      "label": "Measurement",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "recurrent neural networks",
+      "label": "ModelArchitecture",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "contextual representations",
+      "label": "NeuralRepresentation",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "Measurement",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "patterned connectivity perturbations",
+      "label": "Method",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "Method",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "Measurement",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "task variables",
+      "label": "ExperimentalVariable",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "neural response data",
+      "label": "DataType",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "Cognitive functions",
+      "label": "CognitiveFunction",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "sensory",
+      "label": "SignalType",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "contextual signals",
+      "label": "SignalType",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral response",
+      "label": "Measurement",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "excitatory",
+      "label": "CellType",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "inhibitory neurons",
+      "label": "CellType",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cortical circuits",
+      "label": "NeuralCircuit",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "Method",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "excitation",
+      "label": "Phenomenon",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "inhibition",
+      "label": "Phenomenon",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural populations",
+      "label": "CellType",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "task variables",
+      "label": "ExperimentalVariable",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output1–11.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "excitation–inhibition balance",
+      "label": "Phenomenon",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation–inhibition balance12) and thus can be causally validated in experiments13–15.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of single-neuron responses in the cortex.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cortex",
+      "label": "BrainRegion",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of single-neuron responses in the cortex.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "Single neurons",
+      "label": "CellType",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "task variables",
+      "label": "ExperimentalVariable",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables16–21, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "heterogeneous responses",
+      "label": "Measurement",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "recurrent neural network",
+      "label": "ModelArchitecture",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNN",
+      "label": "ModelArchitecture",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural activity data",
+      "label": "DataType",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "recurrent connectivity parameters",
+      "label": "Measurement",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks30.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "Method",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks30.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "fixed points",
+      "label": "Phenomenon",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural response data",
+      "label": "DataType",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs19,31 or fitting a dynamical system directly to neural response data27,32–35.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas30,36.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "population activity",
+      "label": "Measurement",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas30,36.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas30,36.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "brain areas",
+      "label": "BrainRegion",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas30,36.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent variables",
+      "label": "ComputationalEntity",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural population activity",
+      "label": "Measurement",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "unsupervised methods",
+      "label": "Method",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution26,37–40.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "task inputs",
+      "label": "ExperimentalVariable",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution26,37–40.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral outputs",
+      "label": "Measurement",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution26,37–40.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent variables",
+      "label": "ComputationalEntity",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution26,37–40.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural population activity",
+      "label": "Measurement",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "external task variables",
+      "label": "ExperimentalVariable",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables19,21,41–43 (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "However, unlike RNNs and circuit models, these correlation-based methods do not incorporate recurrent interactions among task variables, which implement computations necessary to solve the task.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "task variables",
+      "label": "ExperimentalVariable",
+      "sentence": "However, unlike RNNs and circuit models, these correlation-based methods do not incorporate recurrent interactions among task variables, which implement computations necessary to solve the task.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC’s role in context-dependent decision-making.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "Context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus–response mappings.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "stimulus–response mappings",
+      "label": "ExperimentalParadigm",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus–response mappings.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "NeuralRepresentation",
+      "sentence": "Most circuit models hypothesize a relatively simple mechanism based on inhibition of sensory representations irrelevant in the current context8–10,44–46.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC data",
+      "label": "DataType",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "Measurement",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "These results suggest that RNNs and conceivably the PFC implement qualitatively novel task solutions that do not exist in small circuits and emerge only in high-dimensional networks19,47,48.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "These results suggest that RNNs and conceivably the PFC implement qualitatively novel task solutions that do not exist in small circuits and emerge only in high-dimensional networks19,47,48.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "Method",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "task variables",
+      "label": "ExperimentalVariable",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "heterogeneous neural responses",
+      "label": "Measurement",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "irrelevant sensory representations",
+      "label": "NeuralRepresentation",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods19,21.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "Method",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "single-neuron heterogeneity",
+      "label": "Measurement",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neurons",
+      "label": "CellType",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "cognitive task",
+      "label": "BehavioralAssay",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent variables",
+      "label": "ComputationalEntity",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent nodes",
+      "label": "ComputationalEntity",
+      "sentence": "The latent variables x are constrained to be nodes in a neural circuit with dynamics",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neural circuit",
+      "label": "NeuralCircuit",
+      "sentence": "The latent variables x are constrained to be nodes in a neural circuit with dynamics",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "rectified linear (ReLU) activation function",
+      "label": "ComputationalMethod",
+      "sentence": "where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "ReLU",
+      "label": "ComputationalMethod",
+      "sentence": "where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent nodes",
+      "label": "ComputationalEntity",
+      "sentence": "The latent nodes interact via the recurrent connectivity wrec and receive external task inputs u through the input connectivity win.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "recurrent connectivity",
+      "label": "Measurement",
+      "sentence": "The latent nodes interact via the recurrent connectivity wrec and receive external task inputs u through the input connectivity win.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "external task inputs",
+      "label": "ExperimentalVariable",
+      "sentence": "The latent nodes interact via the recurrent connectivity wrec and receive external task inputs u through the input connectivity win.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "task outputs",
+      "label": "Measurement",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity wout:",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "circuit activity",
+      "label": "Measurement",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity wout:",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "Method",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. (2).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "task-related neural activity",
+      "label": "Measurement",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. (2).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neural circuit",
+      "label": "NeuralCircuit",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. (2).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "task inputs",
+      "label": "ExperimentalVariable",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via win, recurrent interactions among latent nodes via wrec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "cognitive task",
+      "label": "BehavioralAssay",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "mixed selectivity",
+      "label": "Phenomenon",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "cortex",
+      "label": "BrainRegion",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit’s activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "We consider RNNs with dynamics",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "recurrent and input connectivity matrices",
+      "label": "Measurement",
+      "sentence": "where Wrec and Win are the recurrent and input connectivity matrices, respectively (Methods).",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN",
+      "label": "ModelArchitecture",
+      "sentence": "Using the fact that the vector fields of the RNN and latent circuit are piecewise-linear dynamical systems, we derive a relationship between their recurrent and input connectivity matrices (Methods):",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "Using the fact that the vector fields of the RNN and latent circuit are piecewise-linear dynamical systems, we derive a relationship between their recurrent and input connectivity matrices (Methods):",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory stimulus",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context cue",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN",
+      "label": "ModelArchitecture",
+      "sentence": "The RNN successfully learns the task; it makes choices according to the relevant stimulus and ignores the irrelevant stimulus in each context (Fig. 2c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "relevant stimulus",
+      "label": "Stimulus",
+      "sentence": "The RNN successfully learns the task; it makes choices according to the relevant stimulus and ignores the irrelevant stimulus in each context (Fig. 2c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant stimulus",
+      "label": "Stimulus",
+      "sentence": "The RNN successfully learns the task; it makes choices according to the relevant stimulus and ignores the irrelevant stimulus in each context (Fig. 2c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalEntity",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "mixed selectivity",
+      "label": "Phenomenon",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "task variables",
+      "label": "ExperimentalVariable",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "Method",
+      "sentence": "We fitted the latent circuit model to the responses of RNN units and its output behavior during the task.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalEntity",
+      "sentence": "We fitted the latent circuit model to the responses of RNN units and its output behavior during the task.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalEntity",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory color nodes",
+      "label": "ComputationalEntity",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory motion nodes",
+      "label": "ComputationalEntity",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "choice nodes",
+      "label": "ComputationalEntity",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "principal components",
+      "label": "ComputationalEntity",
+      "sentence": "This choice of the latent circuit dimensionality agrees with the dimensionality of RNN responses after training, which is usually close to the total number of inputs and outputs (the first eight principal components accounted for 97.9% of the total variance in RNN responses; Supplementary Fig. 2).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "This choice of the latent circuit dimensionality agrees with the dimensionality of RNN responses after training, which is usually close to the total number of inputs and outputs (the first eight principal components accounted for 97.9% of the total variance in RNN responses; Supplementary Fig. 2).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "left motion",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "green color",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "right motion",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "red color",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color context node",
+      "label": "ComputationalEntity",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalEntity",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion context node",
+      "label": "ComputationalEntity",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalEntity",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant sensory representation",
+      "label": "NeuralRepresentation",
+      "sentence": "Because the irrelevant sensory representation is suppressed, it does not drive the decision output.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "decision output",
+      "label": "Measurement",
+      "sentence": "Because the irrelevant sensory representation is suppressed, it does not drive the decision output.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "First, we verified the signatures of the suppression mechanism in the RNN activity.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "We projected RNN responses onto the columns of Q, which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity patterns",
+      "label": "Measurement",
+      "sentence": "We projected RNN responses onto the columns of Q, which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "We projected RNN responses onto the columns of Q, which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion axis",
+      "label": "ComputationalEntity",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "left and right motion nodes",
+      "label": "ComputationalEntity",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color axis",
+      "label": "ComputationalEntity",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "red and green color nodes",
+      "label": "ComputationalEntity",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "Projections of RNN activity onto the motion and color axes reveal representations of relevant sensory stimuli, whereas representations of irrelevant stimuli are suppressed.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "Measurement",
+      "sentence": "Second, we used the connectivity relationships Eq. (5) to directly validate the latent circuit mechanism in the RNN connectivity.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity matrices",
+      "label": "Measurement",
+      "sentence": "We conjugated the RNN connectivity matrices with the embedding matrix Q.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "embedding matrix Q",
+      "label": "ComputationalEntity",
+      "sentence": "We conjugated the RNN connectivity matrices with the embedding matrix Q.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent connectivity structure",
+      "label": "Measurement",
+      "sentence": "Finally, to ultimately validate that this latent connectivity structure supports the behavioral task performance, we tested whether patterned perturbations of the RNN connectivity (Eq. (6)) produced the same behavioral effects as predicted by the latent circuit model.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "Measurement",
+      "sentence": "Finally, to ultimately validate that this latent connectivity structure supports the behavioral task performance, we tested whether patterned perturbations of the RNN connectivity (Eq. (6)) produced the same behavioral effects as predicted by the latent circuit model.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "Method",
+      "sentence": "Finally, to ultimately validate that this latent connectivity structure supports the behavioral task performance, we tested whether patterned perturbations of the RNN connectivity (Eq. (6)) produced the same behavioral effects as predicted by the latent circuit model.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "We next asked whether RNNs trained on the same task arrive at different circuit solutions for context-dependent decision-making.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "The latent circuit inference enables us to determine whether different RNNs use the same circuit mechanism.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "Although two RNNs trained on the same task may have distinct high-dimensional connectivity, the latent circuit inference can reveal whether these RNNs use similar low-dimensional connectivity structure to generate task-relevant dynamics.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "Although two RNNs trained on the same task may have distinct high-dimensional connectivity, the latent circuit inference can reveal whether these RNNs use similar low-dimensional connectivity structure to generate task-relevant dynamics.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "low-dimensional connectivity structure",
+      "label": "Measurement",
+      "sentence": "Although two RNNs trained on the same task may have distinct high-dimensional connectivity, the latent circuit inference can reveal whether these RNNs use similar low-dimensional connectivity structure to generate task-relevant dynamics.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "task-relevant dynamics",
+      "label": "Measurement",
+      "sentence": "Although two RNNs trained on the same task may have distinct high-dimensional connectivity, the latent circuit inference can reveal whether these RNNs use similar low-dimensional connectivity structure to generate task-relevant dynamics.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "200 RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "To explore the space of circuit mechanisms, we trained an ensemble of 200 RNNs with randomly initialized connectivity to the same level of task performance (r2 = 0.93 ± 0.01, coefficient of determination for the RNN and target outputs, mean ± s.d. across networks; Supplementary Fig. 4).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "principal component analysis",
+      "label": "Method",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices wrec and projected the data onto the first two principal components, which accounted for 42% of total variance (Fig. 5a).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "principal components",
+      "label": "ComputationalEntity",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices wrec and projected the data onto the first two principal components, which accounted for 42% of total variance (Fig. 5a).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "Gaussian mixture model",
+      "label": "Method",
+      "sentence": "The projection reveals three major solution clusters, which we partitioned by fitting data with a Gaussian mixture model with three components.",
+      "paper_location": "Fig. 5 caption"
+    },
+    {
+      "entity": "Mann–Whitney U-test",
+      "label": "Method",
+      "sentence": "These latent models produced a significantly worse fit when the latent circuit and the target RNN were sampled from different clusters than from the same cluster (Extended Data Fig. 2m; one-sided Mann–Whitney U-test, U = 52, 593, P < 10−10), confirming differences in circuit mechanisms across clusters.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Representations of irrelevant stimuli in the PFC",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "Measurement",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed19,21,51.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "We therefore sought to determine whether the latent circuit and regression models identify different task representations in the same PFC responses.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "Method",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC recordings",
+      "label": "DataType",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC neurons",
+      "label": "CellType",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkeys A and F",
+      "label": "Species",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "rhesus monkeys",
+      "label": "Species",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task19 (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting 100 ms after the stimulus onset19 (Methods).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "stimulus onset",
+      "label": "StimulusTiming",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting 100 ms after the stimulus onset19 (Methods).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkeys A and F",
+      "label": "Species",
+      "sentence": "The latent circuit model provided good fits of PFC responses projected onto the low-dimensional subspace spanned by the columns of the inferred embedding matrix Q for both monkeys (r2 = 0.88 and 0.76 on test data for monkeys A and F, respectively; Extended Data Figs. 7a and 8a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "stimulus representations",
+      "label": "NeuralRepresentation",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "Mann–Whitney U-test",
+      "label": "Method",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkey A",
+      "label": "Species",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkey F",
+      "label": "Species",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann–Whitney U-test; monkey A: motion P = 1.4 × 10−9 and U = 120, color P = 10−10 and U = 83; monkey F: motion P = 1.5 × 10−5 and U = 277, color P = 1.6 × 10−7 and U = 194; n = 36).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC activity",
+      "label": "Measurement",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalEntity",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalEntity",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "error trials",
+      "label": "BehavioralAssay",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "Specifically, the suppression mechanism predicts that the representations of irrelevant stimuli (along the motion and color axes identified by the latent circuit model) should be less suppressed on error than correct trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "motion and color axes",
+      "label": "ComputationalEntity",
+      "sentence": "Specifically, the suppression mechanism predicts that the representations of irrelevant stimuli (along the motion and color axes identified by the latent circuit model) should be less suppressed on error than correct trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "Method",
+      "sentence": "Specifically, the suppression mechanism predicts that the representations of irrelevant stimuli (along the motion and color axes identified by the latent circuit model) should be less suppressed on error than correct trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "incongruent trials",
+      "label": "BehavioralAssay",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "relevant and irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "attention lapse",
+      "label": "Phenotype",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "Mann–Whitney U-test",
+      "label": "Method",
+      "sentence": "The prediction was clearly borne out by the data; in both contexts for both monkeys, the irrelevant representations were significantly less suppressed on error than correct trials (Extended Data Figs. 7d and 8d; combined condition Mann–Whitney U-test; monkey A: P = 0.0086, U = 0.313 and n = 24 on color context trials and P = 0.0002, U = 0.354 and n = 24 on motion context trials; monkey F: P = 0.0002, U = 0.333 and n = 26 on color context trials and P < 10−10, U = 0.375 and n = 30 on motion context trials), suggesting that the representations identified by the latent circuit model in PFC activity are related to the behavioral task execution.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkey A",
+      "label": "Species",
+      "sentence": "The prediction was clearly borne out by the data; in both contexts for both monkeys, the irrelevant representations were significantly less suppressed on error than correct trials (Extended Data Figs. 7d and 8d; combined condition Mann–Whitney U-test; monkey A: P = 0.0086, U = 0.313 and n = 24 on color context trials and P = 0.0002, U = 0.354 and n = 24 on motion context trials; monkey F: P = 0.0002, U = 0.333 and n = 26 on color context trials and P < 10−10, U = 0.375 and n = 30 on motion context trials), suggesting that the representations identified by the latent circuit model in PFC activity are related to the behavioral task execution.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkey F",
+      "label": "Species",
+      "sentence": "The prediction was clearly borne out by the data; in both contexts for both monkeys, the irrelevant representations were significantly less suppressed on error than correct trials (Extended Data Figs. 7d and 8d; combined condition Mann–Whitney U-test; monkey A: P = 0.0086, U = 0.313 and n = 24 on color context trials and P = 0.0002, U = 0.354 and n = 24 on motion context trials; monkey F: P = 0.0002, U = 0.333 and n = 26 on color context trials and P < 10−10, U = 0.375 and n = 30 on motion context trials), suggesting that the representations identified by the latent circuit model in PFC activity are related to the behavioral task execution.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "Method",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "Method",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "decoder axis",
+      "label": "ComputationalEntity",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "irrelevant motion stimulus",
+      "label": "Stimulus",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "Single neurons",
+      "label": "CellType",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "Method",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "single-neuron heterogeneity",
+      "label": "Measurement",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "latent variables",
+      "label": "ComputationalEntity",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "high-dimensional networks",
+      "label": "ComputationalModel",
+      "sentence": "We show that low-dimensional circuit mechanisms can explain task-relevant dynamics in high-dimensional networks and establish feasibility of inferring these mechanisms from neural response data.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural response data",
+      "label": "DataType",
+      "sentence": "We show that low-dimensional circuit mechanisms can explain task-relevant dynamics in high-dimensional networks and establish feasibility of inferring these mechanisms from neural response data.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "Our theory for interpreting the latent circuit connectivity as a low-rank connectivity in RNNs enables causally validating low-dimensional circuit mechanisms via activity and connectivity perturbations in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "activity and connectivity perturbations",
+      "label": "Method",
+      "sentence": "Our theory for interpreting the latent circuit connectivity as a low-rank connectivity in RNNs enables causally validating low-dimensional circuit mechanisms via activity and connectivity perturbations in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural data",
+      "label": "DataType",
+      "sentence": "Most methods fitting low-dimensional dynamical systems to neural data do not explicitly model task inputs and behavior and do not ground the inferred dynamics in the underlying network connectivity32,33,40,52,53.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "task inputs",
+      "label": "ExperimentalVariable",
+      "sentence": "Most methods fitting low-dimensional dynamical systems to neural data do not explicitly model task inputs and behavior and do not ground the inferred dynamics in the underlying network connectivity32,33,40,52,53.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "network connectivity",
+      "label": "Measurement",
+      "sentence": "Most methods fitting low-dimensional dynamical systems to neural data do not explicitly model task inputs and behavior and do not ground the inferred dynamics in the underlying network connectivity32,33,40,52,53.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalEntity",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "irrelevant sensory nodes",
+      "label": "ComputationalEntity",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "irrelevant stimulus representations",
+      "label": "NeuralRepresentation",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "We fit the latent circuit model Eqs. (1)–(3) to neural response data y by minimizing the mean squared error loss function, using custom Python code59.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "neural response data",
+      "label": "DataType",
+      "sentence": "We fit the latent circuit model Eqs. (1)–(3) to neural response data y by minimizing the mean squared error loss function, using custom Python code59.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "mean squared error loss function",
+      "label": "Method",
+      "sentence": "We fit the latent circuit model Eqs. (1)–(3) to neural response data y by minimizing the mean squared error loss function, using custom Python code59.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "orthonormal embedding matrix",
+      "label": "ComputationalEntity",
+      "sentence": "Here, k indexes the trials, t indexes the time within a trial, and Q is an orthonormal embedding matrix.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "nonlinear least squares optimization",
+      "label": "Method",
+      "sentence": "Because the variable x depends implicitly on the latent circuit parameters wrec and win, the minimization of L is a nonlinear least squares optimization problem60 in which we simultaneously search for a behaviorally relevant projection of the high-dimensional activity and a low-dimensional neural circuit that generates dynamics in this projection.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "Cayley transform",
+      "label": "Method",
+      "sentence": "To transform it into an unconstrained problem, we use the Cayley transform to parameterize orthonormal matrices by the linear space of skew symmetric matrices61,",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "We perform this minimization using PyTorch and the Adam optimizer with default values 0.9 and 0.999 for the decay rate of the first and second moment estimates, respectively, a learning rate of 0.02 and a weight decay of 0.001.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "Adam optimizer",
+      "label": "Software",
+      "sentence": "We perform this minimization using PyTorch and the Adam optimizer with default values 0.9 and 0.999 for the decay rate of the first and second moment estimates, respectively, a learning rate of 0.02 and a weight decay of 0.001.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "Seaborn",
+      "label": "Software",
+      "sentence": "We used the Python software package Seaborn for visualizing model parameters and responses after training.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "We used the Python software package Seaborn for visualizing model parameters and responses after training.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "Mann–Whitney U-test",
+      "label": "Method",
+      "sentence": "Third, we use a Mann–Whitney U-test to determine whether the correlation coefficients are significantly smaller for models fitted to the shuffled responses than original neural responses.",
+      "paper_location": "Methods: Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "ReLU",
+      "label": "ComputationalMethod",
+      "sentence": "Here, [⋅]+ is a rectified linear (ReLU) activation function, τ is a time constant, and u are external task inputs.",
+      "paper_location": "Methods: Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "external task inputs",
+      "label": "ExperimentalVariable",
+      "sentence": "Here, [⋅]+ is a rectified linear (ReLU) activation function, τ is a time constant, and u are external task inputs.",
+      "paper_location": "Methods: Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "Softplus activation function",
+      "label": "ComputationalMethod",
+      "sentence": "To test whether our results extend to networks with a different biologically plausible nonlinearity, we trained RNNs that had a Softplus activation function f(x) = g β log(1 + eβx) for a range of parameter β and also with varying gain g across units.",
+      "paper_location": "Methods: Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "To test whether our results extend to networks with a different biologically plausible nonlinearity, we trained RNNs that had a Softplus activation function f(x) = g β log(1 + eβx) for a range of parameter β and also with varying gain g across units.",
+      "paper_location": "Methods: Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "We simulate dynamics of time-discretized RNNs using the general framework for modeling cognitive tasks22.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "We simulate dynamics of time-discretized RNNs using the general framework for modeling cognitive tasks22.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "We consider RNNs with positive activity and N = 50 recurrent units.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "recurrent units",
+      "label": "ComputationalEntity",
+      "sentence": "We consider RNNs with positive activity and N = 50 recurrent units.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "Dale’s law",
+      "label": "Phenomenon",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale’s law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "excitatory units",
+      "label": "ComputationalEntity",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale’s law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "inhibitory units",
+      "label": "ComputationalEntity",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale’s law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "The RNN simulation and training were implemented in Python using the software package PyTorch.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "The RNN simulation and training were implemented in Python using the software package PyTorch.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "context cue",
+      "label": "Stimulus",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "color",
+      "label": "StimulusFeature",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion context",
+      "label": "ExperimentalCondition",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "sensory stimulus",
+      "label": "Stimulus",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion",
+      "label": "StimulusFeature",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "color features",
+      "label": "StimulusFeature",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "right motion",
+      "label": "Stimulus",
+      "sentence": "The right motion and red color are associated with the right choice, and the left motion and green color are associated with the left choice.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "red color",
+      "label": "Stimulus",
+      "sentence": "The right motion and red color are associated with the right choice, and the left motion and green color are associated with the left choice.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "left motion",
+      "label": "Stimulus",
+      "sentence": "The right motion and red color are associated with the right choice, and the left motion and green color are associated with the left choice.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "green color",
+      "label": "Stimulus",
+      "sentence": "The right motion and red color are associated with the right choice, and the left motion and green color are associated with the left choice.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "color coherence",
+      "label": "Measurement",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "color context",
+      "label": "ExperimentalCondition",
+      "sentence": "In the color context, the choice should be made according to the color, ignoring the motion stimulus, and vice versa in the motion context.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion stimulus",
+      "label": "Stimulus",
+      "sentence": "In the color context, the choice should be made according to the color, ignoring the motion stimulus, and vice versa in the motion context.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion context",
+      "label": "ExperimentalCondition",
+      "sentence": "In the color context, the choice should be made according to the color, ignoring the motion stimulus, and vice versa in the motion context.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "Adam algorithm",
+      "label": "Software",
+      "sentence": "The training is performed with the Adam algorithm.",
+      "paper_location": "Methods: RNN training"
+    },
+    {
+      "entity": "linear regression model",
+      "label": "Method",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model",
+      "paper_location": "Methods: Linear decoding"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model",
+      "paper_location": "Methods: Linear decoding"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model",
+      "paper_location": "Methods: Linear decoding"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "frontal eye field",
+      "label": "BrainRegion",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task19.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "monkey A",
+      "label": "Species",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "monkey F",
+      "label": "Species",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "multiunits",
+      "label": "CellType",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "neural data",
+      "label": "DataType",
+      "sentence": "To facilitate comparison with previous studies analyzing the same dataset19,21, we used identical initial preprocessing of the neural data (using the publicly available code at https://www.ini.uzh.ch/en/research/groups/mante/data.html).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "saccade",
+      "label": "BehavioralResponse",
+      "sentence": "Monkeys reported their choice with a saccade to one of two targets presented shortly after fixation for the entire trial duration.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "fixation",
+      "label": "BehavioralAssay",
+      "sentence": "Monkeys reported their choice with a saccade to one of two targets presented shortly after fixation for the entire trial duration.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "saccades",
+      "label": "BehavioralResponse",
+      "sentence": "The monkeys were rewarded for saccades to the target location corresponding to the motion direction in the motion context and to the target whose color matched the dominant color of the dots in the color context.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "motion direction",
+      "label": "StimulusFeature",
+      "sentence": "The monkeys were rewarded for saccades to the target location corresponding to the motion direction in the motion context and to the target whose color matched the dominant color of the dots in the color context.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "motion context",
+      "label": "ExperimentalCondition",
+      "sentence": "The monkeys were rewarded for saccades to the target location corresponding to the motion direction in the motion context and to the target whose color matched the dominant color of the dots in the color context.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "dominant color of the dots",
+      "label": "Stimulus",
+      "sentence": "The monkeys were rewarded for saccades to the target location corresponding to the motion direction in the motion context and to the target whose color matched the dominant color of the dots in the color context.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "color context",
+      "label": "ExperimentalCondition",
+      "sentence": "The monkeys were rewarded for saccades to the target location corresponding to the motion direction in the motion context and to the target whose color matched the dominant color of the dots in the color context.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "PFC data",
+      "label": "DataType",
+      "sentence": "We fitted latent circuit models to the PFC data following similar procedures as for RNNs.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ModelArchitecture",
+      "sentence": "We fitted latent circuit models to the PFC data following similar procedures as for RNNs.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "Gaussian kernel",
+      "label": "Method",
+      "sentence": "Within the training and test sets, we z scored and smoothed (Gaussian kernel, σ = 40 ms) the response of each unit.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "principal components",
+      "label": "ComputationalEntity",
+      "sentence": "Last, to denoise these trajectories, we projected them onto the principal components explaining 50% of their total variance (corresponding to the first 40 and 31 principal components for monkeys A and F, respectively).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "principal components",
+      "label": "ComputationalEntity",
+      "sentence": "Last, to denoise these trajectories, we projected them onto the principal components explaining 50% of their total variance (corresponding to the first 40 and 31 principal components for monkeys A and F, respectively).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "monkeys A and F",
+      "label": "Species",
+      "sentence": "Last, to denoise these trajectories, we projected them onto the principal components explaining 50% of their total variance (corresponding to the first 40 and 31 principal components for monkeys A and F, respectively).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "200 RNN models",
+      "label": "ModelArchitecture",
+      "sentence": "We analyzed data from 200 RNN models trained with random initializations.",
+      "paper_location": "Methods: Statistics and reproducibility"
+    },
+    {
+      "entity": "100 latent circuit models",
+      "label": "ComputationalModel",
+      "sentence": "For each of these networks, we trained 100 latent circuit models.",
+      "paper_location": "Methods: Statistics and reproducibility"
+    },
+    {
+      "entity": "PFC data",
+      "label": "DataType",
+      "sentence": "For PFC data, we fitted 200 latent circuit models to neural responses from each monkey.",
+      "paper_location": "Methods: Statistics and reproducibility"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "For PFC data, we fitted 200 latent circuit models to neural responses from each monkey.",
+      "paper_location": "Methods: Statistics and reproducibility"
+    },
+    {
+      "entity": "monkey",
+      "label": "Species",
+      "sentence": "For PFC data, we fitted 200 latent circuit models to neural responses from each monkey.",
+      "paper_location": "Methods: Statistics and reproducibility"
+    },
+    {
+      "entity": "Neural recording data",
+      "label": "DataType",
+      "sentence": "Neural recording data were previously described in Mante et al.19; no randomization or blinding was performed because there was only one experimental group.",
+      "paper_location": "Methods: Statistics and reproducibility"
+    },
+    {
+      "entity": "Synthetic data",
+      "label": "DataType",
+      "sentence": "Synthetic data used in this study can be reproduced using the source code with model parameters included in the Source Data files provided with this paper.",
+      "paper_location": "Data availability"
+    },
+    {
+      "entity": "Neural recording data",
+      "label": "DataType",
+      "sentence": "Neural recording data were previously described in Mante et al.19 and are available in a public repository at https://www.ini.uzh.ch/en/research/groups/mante/data.html.",
+      "paper_location": "Data availability"
+    },
+    {
+      "entity": "GitHub",
+      "label": "SoftwareRepository",
+      "sentence": "The source code to reproduce the results of this study is publicly available on GitHub (https://github.com/engellab/latentcircuit) and Zenodo (https://zenodo.org/records/14020108)59.",
+      "paper_location": "Code availability"
+    },
+    {
+      "entity": "Zenodo",
+      "label": "SoftwareRepository",
+      "sentence": "The source code to reproduce the results of this study is publicly available on GitHub (https://github.com/engellab/latentcircuit) and Zenodo (https://zenodo.org/records/14020108)59.",
+      "paper_location": "Code availability"
+    },
+    {
+      "entity": "National Institutes of Health",
+      "label": "Organization",
+      "sentence": "This work was supported by National Institutes of Health grants RF1DA055666 and S10OD028632-01 (T.A.E.), the Alfred P. Sloan Foundation Research Fellowship (T.A.E.) and the Swartz Foundation Postdoctoral Fellowship (C.L.).",
+      "paper_location": "Acknowledgements"
+    },
+    {
+      "entity": "RF1DA055666",
+      "label": "Grant",
+      "sentence": "This work was supported by National Institutes of Health grants RF1DA055666 and S10OD028632-01 (T.A.E.), the Alfred P. Sloan Foundation Research Fellowship (T.A.E.) and the Swartz Foundation Postdoctoral Fellowship (C.L.).",
+      "paper_location": "Acknowledgements"
+    },
+    {
+      "entity": "S10OD028632-01",
+      "label": "Grant",
+      "sentence": "This work was supported by National Institutes of Health grants RF1DA055666 and S10OD028632-01 (T.A.E.), the Alfred P. Sloan Foundation Research Fellowship (T.A.E.) and the Swartz Foundation Postdoctoral Fellowship (C.L.).",
+      "paper_location": "Acknowledgements"
+    },
+    {
+      "entity": "MAP data-acquisition system",
+      "label": "Instrument",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "Plexon Inc.",
+      "label": "Organization",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "Spike sorting",
+      "label": "Method",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "principal component analysis",
+      "label": "Method",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "Plexon offline sorter",
+      "label": "Software",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "Custom Python code was used for data analyses.",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "Plots were generated with Python (version 3.8) and Seaborn (version 0.9.0).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "Seaborn",
+      "label": "Software",
+      "sentence": "Plots were generated with Python (version 3.8) and Seaborn (version 0.9.0).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "Recurrent neural networks were trained using PyTorch (version 2.4).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "Recurrent neural networks",
+      "label": "ModelArchitecture",
+      "sentence": "Recurrent neural networks were trained using PyTorch (version 2.4).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "adult males",
+      "label": "OrganismAttribute",
+      "sentence": "The animals were adult males, ages 13 (monkey A) and 8 (monkey F) years old at the time the experiments were conducted.",
+      "paper_location": "Reporting Summary: Animals and other research organisms"
+    },
+    {
+      "entity": "monkey A",
+      "label": "Species",
+      "sentence": "The animals were adult males, ages 13 (monkey A) and 8 (monkey F) years old at the time the experiments were conducted.",
+      "paper_location": "Reporting Summary: Animals and other research organisms"
+    },
+    {
+      "entity": "monkey F",
+      "label": "Species",
+      "sentence": "The animals were adult males, ages 13 (monkey A) and 8 (monkey F) years old at the time the experiments were conducted.",
+      "paper_location": "Reporting Summary: Animals and other research organisms"
+    },
+    {
+      "entity": "National Institutes of Health",
+      "label": "Organization",
+      "sentence": "All surgical and behavioral procedures conformed to the guidelines established by the National Institutes of Health and were approved by the Institutional Animal Care and Use Committee of Stanford University.",
+      "paper_location": "Reporting Summary: Animals and other research organisms"
+    },
+    {
+      "entity": "Institutional Animal Care and Use Committee of Stanford University",
+      "label": "Organization",
+      "sentence": "All surgical and behavioral procedures conformed to the guidelines established by the National Institutes of Health and were approved by the Institutional Animal Care and Use Committee of Stanford University.",
+      "paper_location": "Reporting Summary: Animals and other research organisms"
+    }
+  ],
+  "key_terms": [
+    {
+      "term": "latent circuit inference",
+      "sentence": "Latent circuit inference from heterogeneous neural responses during cognitive tasks",
+      "paper_location": "Title"
+    },
+    {
+      "term": "heterogeneous neural responses",
+      "sentence": "Latent circuit inference from heterogeneous neural responses during cognitive tasks",
+      "paper_location": "Title"
+    },
+    {
+      "term": "cognitive tasks",
+      "sentence": "Latent circuit inference from heterogeneous neural responses during cognitive tasks",
+      "paper_location": "Title"
+    },
+    {
+      "term": "dimensionality reduction approach",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "context-dependent decision-making task",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "suppression mechanism",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "causal interactions among task variables",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "behaviorally relevant computations",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "circuit mechanism",
+      "sentence": "Thus, these models explicitly specify a circuit mechanism in the connectivity structure that gives rise to a dynamical mechanism controlling the flow of neural trajectories to implement task computations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "dynamical mechanism",
+      "sentence": "Thus, these models explicitly specify a circuit mechanism in the connectivity structure that gives rise to a dynamical mechanism controlling the flow of neural trajectories to implement task computations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "single-neuron responses",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of single-neuron responses in the cortex.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "high-dimensional recurrent neural network (RNN) models",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks19,22–24 or reproduce neural activity data25–29 by optimizing recurrent connectivity parameters.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "correlation-based methods",
+      "sentence": "However, unlike RNNs and circuit models, these correlation-based methods do not incorporate recurrent interactions among task variables, which implement computations necessary to solve the task.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "context-dependent decision-making",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC’s role in context-dependent decision-making.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "alternative stimulus–response mappings",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus–response mappings.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "inhibitory circuit mechanism",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses19,21, seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "low-dimensional circuit mechanism",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "causal mechanism of task execution",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "low-dimensional latent circuit connectivity",
+      "sentence": "The model simultaneously infers a low-dimensional latent circuit connectivity generating task-relevant dynamics and heterogeneous mixing of these dynamics in single-neuron responses.",
+      "paper_location": "Introduction"
+    },
+    {
+      "term": "low-dimensional latent variables",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝN (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝn as",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "term": "orthonormal embedding matrix",
+      "sentence": "where Q ∈ ℝN×n is an orthonormal embedding matrix and n ≪ N.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "term": "rectified linear (ReLU) activation function",
+      "sentence": "where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "term": "recurrent connectivity wrec",
+      "sentence": "The latent nodes interact via the recurrent connectivity wrec and receive external task inputs u through the input connectivity win.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "term": "embedding Eq. (1)",
+      "sentence": "To interpret the latent connectivity, we differentiate the embedding Eq. (1) to obtain the correspondence between vector fields of the high-dimensional and low-dimensional dynamical systems (Fig. 1d and Methods).",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "term": "piecewise-linear dynamical systems",
+      "sentence": "Using the fact that the vector fields of the RNN and latent circuit are piecewise-linear dynamical systems, we derive a relationship between their recurrent and input connectivity matrices (Methods):",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "term": "rank-one connectivity perturbation",
+      "sentence": "Perturbing connection δij from node j to node i in the latent circuit maps onto rank-one connectivity perturbation QδijQT = qiqT j in the RNN.",
+      "paper_location": "Fig. 1 caption"
+    },
+    {
+      "term": "context cue",
+      "sentence": "Each trial begins with a brief baseline period (hold).",
+      "paper_location": "Fig. 2 caption"
+    },
+    {
+      "term": "sensory stimulus",
+      "sentence": "After a short delay (delay), a sensory stimulus appears that consists of motion and color features (stimulus), and a response can be made at any time.",
+      "paper_location": "Fig. 2 caption"
+    },
+    {
+      "term": "psychometric functions",
+      "sentence": "Psychometric functions show that the RNN successfully learns the task; it responds to relevant stimuli and ignores irrelevant stimuli in each context.",
+      "paper_location": "Fig. 2 caption"
+    },
+    {
+      "term": "inhibition of irrelevant representations",
+      "sentence": "This suppression mechanism based on inhibition of irrelevant representations is qualitatively similar to mechanisms for context-dependent decision-making hypothesized in previous hand-crafted neural circuit models8,9.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "term": "rank-one patterned connectivity perturbation",
+      "sentence": "The perturbation in a of the latent circuit connectivity maps onto rank-one patterned connectivity perturbation in the RNN (left).",
+      "paper_location": "Fig. 4 caption"
+    },
+    {
+      "term": "space of circuit mechanisms",
+      "sentence": "To explore the space of circuit mechanisms, we trained an ensemble of 200 RNNs with randomly initialized connectivity to the same level of task performance (r2 = 0.93 ± 0.01, coefficient of determination for the RNN and target outputs, mean ± s.d. across networks; Supplementary Fig. 4).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "term": "converged solutions",
+      "sentence": "For each of these RNNs, we fitted an ensemble of 100 latent circuit models starting with random initializations of the parameters for latent connectivity and embedding Q and selected 10 latent circuits with the highest fit quality on the test data, which formed the set of converged solutions (Methods).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "term": "asymmetric inhibitory mechanism",
+      "sentence": "Asymmetric inhibitory mechanism for context-dependent decision-making.",
+      "paper_location": "Extended Data Fig. 3 caption"
+    },
+    {
+      "term": "dynamical selection vector mechanism",
+      "sentence": "Dynamical selection vector mechanism in RNNs and latent circuits that implement the inhibitory suppression mechanism.",
+      "paper_location": "Extended Data Fig. 4 caption"
+    },
+    {
+      "term": "line attractor",
+      "sentence": "In this mechanism, choice is integrated along a line attractor in each context (l, choice axis).",
+      "paper_location": "Extended Data Fig. 4 caption"
+    },
+    {
+      "term": "selection vector",
+      "sentence": "Only inputs aligned with the left eigenvector (vL, selection vector) drive activity along the right eigenvector (vR), which has a non-zero projection onto the choice axis l.",
+      "paper_location": "Extended Data Fig. 4 caption"
+    },
+    {
+      "term": "task subspace",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts19,21.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "term": "regression models",
+      "sentence": "To confirm that this difference results from recurrent interactions among task variables in the latent circuit model, we fitted PFC data with a modified latent circuit model in which the latent recurrent connectivity was constrained to be 0.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "term": "error trials",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "term": "behavioral relevance of low-dimensional representations",
+      "sentence": "Behavioral relevance of low-dimensional representations",
+      "paper_location": "Results heading"
+    },
+    {
+      "term": "decoder axis",
+      "sentence": "The decoding weights provide an axis in the RNN state space such that a projection onto this axis correlates with the motion coherence.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "term": "demixing representations of task variables",
+      "sentence": "Our results indicate that ‘demixing’ representations of task variables19,21,42 may not be the right objective for identifying behaviorally relevant patterns in neural activity and may provide a misleading picture of computation.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "term": "low-rank connectivity",
+      "sentence": "Our theory for interpreting the latent circuit connectivity as a low-rank connectivity in RNNs enables causally validating low-dimensional circuit mechanisms via activity and connectivity perturbations in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "latent dynamics generators",
+      "sentence": "Our results motivate future work incorporating low-dimensional recurrent circuits as latent dynamics generators within these model architectures, which will enable combining their ability of fitting single-trial neural activity26,29 with the uniqueness and interpretability furnished by the latent circuit model.",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "cognitive flexibility",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "mean squared error loss function",
+      "sentence": "We fit the latent circuit model Eqs. (1)–(3) to neural response data y by minimizing the mean squared error loss function, using custom Python code59.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "term": "nonconvex optimization",
+      "sentence": "This ensemble of latent circuit models for a single RNN has variable fit quality because many optimization runs do not converge to the optimal solution (which is typical for nonconvex optimization).",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "term": "permutation test",
+      "sentence": "To determine whether the inferred latent circuit connectivity significantly depends on neural response data beyond the constraints imposed by the task alone, we performed a permutation test (Extended Data Figs. 1, 7c and 8c), which proceeds in three steps.",
+      "paper_location": "Methods: Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "term": "time-discretized RNNs",
+      "sentence": "We simulate dynamics of time-discretized RNNs using the general framework for modeling cognitive tasks22.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "term": "Dale’s law",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale’s law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "term": "RNN training",
+      "sentence": "RNN training",
+      "paper_location": "Methods heading"
+    },
+    {
+      "term": "Linear decoding",
+      "sentence": "Linear decoding",
+      "paper_location": "Methods heading"
+    },
+    {
+      "term": "Analysis of PFC data",
+      "sentence": "Analysis of PFC data",
+      "paper_location": "Methods heading"
+    },
+    {
+      "term": "spike sorting",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_openai-gpt-5.5_20260629_095709.json b/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_openai-gpt-5.5_20260629_095709.json
new file mode 100644
index 0000000..1d5be21
--- /dev/null
+++ b/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_openai-gpt-5.5_20260629_095709.json
@@ -0,0 +1,3221 @@
+{
+  "source_metadata": {
+    "source_path": "evaluation/ner/latent-circuit-inference/s41593-025-01869-7.pdf",
+    "statistics": {
+      "model": "openai/gpt-5.5",
+      "prompt_file": "/Users/pujatrivedi/Desktop/MIT/structsense/evaluation/ner/direct_llm_call/prompts/extractor_neuroscience_ner.txt",
+      "pdf_engine": "mistral-ocr",
+      "temperature": null,
+      "seed": null,
+      "finish_reason": "stop",
+      "extracted_at": "20260629_095709",
+      "total_entities": 492,
+      "entities_by_label": {
+        "Measurement": 110,
+        "ComputationalModel": 75,
+        "Phenomenon": 41,
+        "Method": 34,
+        "Stimulus": 34,
+        "NeuralCircuit": 29,
+        "BehavioralAssay": 27,
+        "BrainRegion": 22,
+        "ComputationalUnit": 20,
+        "CellType": 16,
+        "ExperimentalCondition": 14,
+        "Species": 11,
+        "Software": 11,
+        "ActivationFunction": 6,
+        "NeuralSignal": 5,
+        "AnimalSubject": 5,
+        "ExperimentalVariable": 4,
+        "TemporalParameter": 4,
+        "CognitiveProcess": 3,
+        "StatisticalTest": 3,
+        "StatisticalMethod": 3,
+        "BehavioralParadigm": 2,
+        "BiologicalPrinciple": 2,
+        "SoftwareRepository": 2,
+        "ModelParameter": 1,
+        "Phenotype": 1,
+        "MathematicalMethod": 1,
+        "NumericalMethod": 1,
+        "BehavioralResponse": 1,
+        "Instrument": 1,
+        "Company": 1,
+        "Organization": 1,
+        "EthicsCommittee": 1
+      }
+    }
+  },
+  "entities": [
+    {
+      "entity": "Latent circuit inference",
+      "label": "Method",
+      "sentence": "Latent circuit inference from heterogeneous neural responses during cognitive tasks",
+      "paper_location": "Title"
+    },
+    {
+      "entity": "heterogeneous neural responses",
+      "label": "Phenomenon",
+      "sentence": "Latent circuit inference from heterogeneous neural responses during cognitive tasks",
+      "paper_location": "Title"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Latent circuit inference from heterogeneous neural responses during cognitive tasks",
+      "paper_location": "Title"
+    },
+    {
+      "entity": "Higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "sensory",
+      "label": "NeuralSignal",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "cognitive",
+      "label": "NeuralSignal",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "motor signals",
+      "label": "NeuralSignal",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "Dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "heterogeneous responses",
+      "label": "Phenomenon",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "dimensionality reduction approach",
+      "label": "Method",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "recurrent neural networks",
+      "label": "ComputationalModel",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "contextual representations",
+      "label": "Phenomenon",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "Measurement",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "behavioral effects",
+      "label": "Measurement",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "connectivity perturbations",
+      "label": "Method",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "Measurement",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "task variables",
+      "label": "ExperimentalVariable",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "entity": "Cognitive functions",
+      "label": "CognitiveProcess",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "sensory",
+      "label": "NeuralSignal",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "contextual signals",
+      "label": "NeuralSignal",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "behavioral response",
+      "label": "Measurement",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "excitatory",
+      "label": "CellType",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "inhibitory neurons",
+      "label": "CellType",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "cortical circuits",
+      "label": "NeuralCircuit",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output^{1--15}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "excitation",
+      "label": "Phenomenon",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output^{1--15}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "inhibition",
+      "label": "Phenomenon",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output^{1--15}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "neural populations",
+      "label": "CellType",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output^{1--15}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output^{1--15}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "circuit models",
+      "label": "ComputationalModel",
+      "sentence": "Because these circuit models usually assume a relatively simple connectivity structure, their connectivity can be directly related to a dynamical system description of computations supporting cognitive task execution^{3,6}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "connectivity structure",
+      "label": "NeuralCircuit",
+      "sentence": "Because these circuit models usually assume a relatively simple connectivity structure, their connectivity can be directly related to a dynamical system description of computations supporting cognitive task execution^{3,6}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "dynamical system",
+      "label": "ComputationalModel",
+      "sentence": "Because these circuit models usually assume a relatively simple connectivity structure, their connectivity can be directly related to a dynamical system description of computations supporting cognitive task execution^{3,6}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "cognitive task execution",
+      "label": "BehavioralAssay",
+      "sentence": "Because these circuit models usually assume a relatively simple connectivity structure, their connectivity can be directly related to a dynamical system description of computations supporting cognitive task execution^{3,6}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "neural trajectories",
+      "label": "Measurement",
+      "sentence": "Thus, these models explicitly specify a circuit mechanism in the connectivity structure that gives rise to a dynamical mechanism controlling the flow of neural trajectories to implement task computations.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "neural dynamics",
+      "label": "Phenomenon",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation--inhibition balance^{12}) and thus can be causally validated in experiments^{13--15}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation--inhibition balance^{12}) and thus can be causally validated in experiments^{13--15}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "behavioral performance",
+      "label": "Measurement",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation--inhibition balance^{12}) and thus can be causally validated in experiments^{13--15}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "excitation--inhibition balance",
+      "label": "Phenomenon",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation--inhibition balance^{12}) and thus can be causally validated in experiments^{13--15}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of single-neuron responses in the cortex.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "cortex",
+      "label": "BrainRegion",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of single-neuron responses in the cortex.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "Single neurons",
+      "label": "CellType",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables^{16--21}, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables^{16--21}, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables^{16--21}, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "heterogeneous tuning",
+      "label": "Phenomenon",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables^{16--21}, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "circuit mechanisms",
+      "label": "NeuralCircuit",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables^{16--21}, which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "heterogeneous responses",
+      "label": "Phenomenon",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks^{19,22--24} or reproduce neural activity data^{25--29} by optimizing recurrent connectivity parameters.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "recurrent neural network",
+      "label": "ComputationalModel",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks^{19,22--24} or reproduce neural activity data^{25--29} by optimizing recurrent connectivity parameters.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks^{19,22--24} or reproduce neural activity data^{25--29} by optimizing recurrent connectivity parameters.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks^{19,22--24} or reproduce neural activity data^{25--29} by optimizing recurrent connectivity parameters.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "neural activity data",
+      "label": "Measurement",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks^{19,22--24} or reproduce neural activity data^{25--29} by optimizing recurrent connectivity parameters.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks^{30}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks^{30}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "fixed points",
+      "label": "Phenomenon",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs^{19,31} or fitting a dynamical system directly to neural response data^{27,32--35}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "linearized dynamics",
+      "label": "Phenomenon",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs^{19,31} or fitting a dynamical system directly to neural response data^{27,32--35}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs^{19,31} or fitting a dynamical system directly to neural response data^{27,32--35}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "dynamical system",
+      "label": "ComputationalModel",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs^{19,31} or fitting a dynamical system directly to neural response data^{27,32--35}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs^{19,31} or fitting a dynamical system directly to neural response data^{27,32--35}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Thus, high-dimensional RNNs currently serve as intermediate mechanistic models of heterogeneous neural responses,",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "heterogeneous neural responses",
+      "label": "Phenomenon",
+      "sentence": "Thus, high-dimensional RNNs currently serve as intermediate mechanistic models of heterogeneous neural responses,",
+      "paper_location": "Abstract"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas $^{30,36}$ .",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "population activity",
+      "label": "Measurement",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas $^{30,36}$ .",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas $^{30,36}$ .",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "brain areas",
+      "label": "BrainRegion",
+      "sentence": "Although single-neuron responses are complex and heterogeneous, their joint population activity is often low dimensional across many cognitive tasks and brain areas $^{30,36}$ .",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural population activity",
+      "label": "Measurement",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "unsupervised methods",
+      "label": "Method",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution $^{26,37-40}$ .",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "task inputs",
+      "label": "ExperimentalVariable",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution $^{26,37-40}$ .",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "behavioral outputs",
+      "label": "Measurement",
+      "sentence": "Because unsupervised methods do not explicitly model task inputs and behavioral outputs, the latent variables they infer may be unrelated to the cognitive task execution $^{26,37-40}$ .",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "targeted dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables $^{19,21,41-45}$  (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural representations",
+      "label": "Phenomenon",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables $^{19,21,41-45}$  (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural population activity",
+      "label": "Measurement",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables $^{19,21,41-45}$  (Fig. 1a,b).",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "However, unlike RNNs and circuit models, these correlation-based methods do not incorporate recurrent interactions among task variables, which implement computations necessary to solve the task.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "circuit models",
+      "label": "ComputationalModel",
+      "sentence": "However, unlike RNNs and circuit models, these correlation-based methods do not incorporate recurrent interactions among task variables, which implement computations necessary to solve the task.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC's role in context-dependent decision-making.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC's role in context-dependent decision-making.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "Context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus-response mappings.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "stimulus-response mappings",
+      "label": "BehavioralParadigm",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus-response mappings.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "Phenomenon",
+      "sentence": "Most circuit models hypothesize a relatively simple mechanism based on inhibition of sensory representations irrelevant in the current context[8-10,44-46].",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses[19,21], seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses[19,21], seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "Measurement",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses[19,21], seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "These results suggest that RNNs and conceivably the PFC implement qualitatively novel task solutions that do not exist in small circuits and emerge only in high-dimensional networks $^{19,47,48}$ .",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "These results suggest that RNNs and conceivably the PFC implement qualitatively novel task solutions that do not exist in small circuits and emerge only in high-dimensional networks $^{19,47,48}$ .",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "task behavior",
+      "label": "Measurement",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "irrelevant sensory representations",
+      "label": "Phenomenon",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We validated this mechanism by confirming the behavioral effects of patterned perturbations of the RNN activity and connectivity predicted by the latent circuit model.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We validated this mechanism by confirming the behavioral effects of patterned perturbations of the RNN activity and connectivity predicted by the latent circuit model.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods[19,21].",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods[19,21].",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods[19,21].",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods[19,21].",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "RNN perturbations",
+      "label": "Method",
+      "sentence": "Using RNN perturbations, we show that dimensionality reduction methods that do not incorporate causal interactions among latent variables are biased toward uncovering behaviorally irrelevant representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Using RNN perturbations, we show that dimensionality reduction methods that do not incorporate causal interactions among latent variables are biased toward uncovering behaviorally irrelevant representations.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "Our results show that high-dimensional networks use low-dimensional circuit mechanisms, establish the feasibility of inferring these mechanisms from neural response data and open new possibilities of causally validating circuit mechanisms in perturbation experiments.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "perturbation experiments",
+      "label": "Method",
+      "sentence": "Our results show that high-dimensional networks use low-dimensional circuit mechanisms, establish the feasibility of inferring these mechanisms from neural response data and open new possibilities of causally validating circuit mechanisms in perturbation experiments.",
+      "paper_location": "Introduction"
+    },
+    {
+      "entity": "Latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "Latent circuit model",
+      "paper_location": "Results"
+    },
+    {
+      "entity": "dimensionality reduction",
+      "label": "Method",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "single-neuron heterogeneity",
+      "label": "Phenomenon",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses *y* ∈ ℝ^{*N*} (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝ^{*n*} as$$y=Qx,$$where Q ∈ ℝ^{*N* × *n*} is an orthonormal embedding matrix and n ≪ N.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses *y* ∈ ℝ^{*N*} (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝ^{*n*} as$$y=Qx,$$where Q ∈ ℝ^{*N* × *n*} is an orthonormal embedding matrix and n ≪ N.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neurons",
+      "label": "CellType",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses *y* ∈ ℝ^{*N*} (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝ^{*n*} as$$y=Qx,$$where Q ∈ ℝ^{*N* × *n*} is an orthonormal embedding matrix and n ≪ N.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "cognitive task",
+      "label": "BehavioralAssay",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses *y* ∈ ℝ^{*N*} (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝ^{*n*} as$$y=Qx,$$where Q ∈ ℝ^{*N* × *n*} is an orthonormal embedding matrix and n ≪ N.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "rectified linear",
+      "label": "ActivationFunction",
+      "sentence": "The latent variables x are constrained to be nodes in a neural circuit with dynamics$$\\dot{x}=-x+f(w_{\\text{rec}}x+w_{\\text{in}}u),$$where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "ReLU",
+      "label": "ActivationFunction",
+      "sentence": "The latent variables x are constrained to be nodes in a neural circuit with dynamics$$\\dot{x}=-x+f(w_{\\text{rec}}x+w_{\\text{in}}u),$$where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "activation function",
+      "label": "ActivationFunction",
+      "sentence": "The latent variables x are constrained to be nodes in a neural circuit with dynamics$$\\dot{x}=-x+f(w_{\\text{rec}}x+w_{\\text{in}}u),$$where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "The latent nodes interact via the recurrent connectivity w_{rec} and receive external task inputs u through the input connectivity w_{in}.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "recurrent connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "The latent nodes interact via the recurrent connectivity w_{rec} and receive external task inputs u through the input connectivity w_{in}.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "task inputs",
+      "label": "ExperimentalVariable",
+      "sentence": "The latent nodes interact via the recurrent connectivity w_{rec} and receive external task inputs u through the input connectivity w_{in}.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity w_{out}:$$z=w_{{\\text{out}}}x.$$",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "task outputs",
+      "label": "Measurement",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity w_{out}:$$z=w_{{\\text{out}}}x.$$",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "circuit activity",
+      "label": "Measurement",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity w_{out}:$$z=w_{{\\text{out}}}x.$$",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. (2).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "task-related neural activity",
+      "label": "Measurement",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. (2).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neural circuit",
+      "label": "NeuralCircuit",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. (2).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent circuit parameters",
+      "label": "ModelParameter",
+      "sentence": "We infer the latent circuit parameters (Q, w_{rec}, w_{in} and w_{out}) from neural activity y by minimizing the loss function *L* = ∑_{*k*, *i*}|*y*-*Q**x*|_{*L*} + ||*z*-*w*_{out}*x*|_{*L*}, where k and t index trials and time within trial, respectively (Methods).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "We infer the latent circuit parameters (Q, w_{rec}, w_{in} and w_{out}) from neural activity y by minimizing the loss function *L* = ∑_{*k*, *i*}|*y*-*Q**x*|_{*L*} + ||*z*-*w*_{out}*x*|_{*L*}, where k and t index trials and time within trial, respectively (Methods).",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via w_{in}, recurrent interactions among latent nodes via w_{rec} and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via w_{in}, recurrent interactions among latent nodes via w_{rec} and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "task inputs",
+      "label": "ExperimentalVariable",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via w_{in}, recurrent interactions among latent nodes via w_{rec} and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via w_{in}, recurrent interactions among latent nodes via w_{rec} and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results: Latent circuit model"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "cognitive task",
+      "label": "BehavioralAssay",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit's activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "mixed selectivity",
+      "label": "Phenomenon",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit's activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit's activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "cortex",
+      "label": "BrainRegion",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit's activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit's activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "network connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit's activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "behavioral outcomes",
+      "label": "Measurement",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit's activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "We can then derive an explicit relationship between connectivity matrices of a high-dimensional RNN and a low-dimensional latent circuit.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "We can then derive an explicit relationship between connectivity matrices of a high-dimensional RNN and a low-dimensional latent circuit.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We consider RNNs with dynamics$$\\dot{y}=-y+f({W}_{\\text{rec}},y+W_{\\text{in}}u),$$where W_{rec} and W_{in} are the recurrent and input connectivity matrices, respectively (Methods).",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "recurrent",
+      "label": "NeuralCircuit",
+      "sentence": "We consider RNNs with dynamics$$\\dot{y}=-y+f({W}_{\\text{rec}},y+W_{\\text{in}}u),$$where W_{rec} and W_{in} are the recurrent and input connectivity matrices, respectively (Methods).",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "input connectivity matrices",
+      "label": "NeuralCircuit",
+      "sentence": "We consider RNNs with dynamics$$\\dot{y}=-y+f({W}_{\\text{rec}},y+W_{\\text{in}}u),$$where W_{rec} and W_{in} are the recurrent and input connectivity matrices, respectively (Methods).",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Using the fact that the vector fields of the RNN and latent circuit are piecewise-linear dynamical systems, we derive a relationship between their recurrent and input connectivity matrices (Methods):$$Q^{T}{W}_{\\text{rec}}Q={w}_{\\text{rec}}, ,Q^{T}{W}_{\\text{in}}={w}_{\\text{in}}.$$",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "Using the fact that the vector fields of the RNN and latent circuit are piecewise-linear dynamical systems, we derive a relationship between their recurrent and input connectivity matrices (Methods):$$Q^{T}{W}_{\\text{rec}}Q={w}_{\\text{rec}}, ,Q^{T}{W}_{\\text{in}}={w}_{\\text{in}}.$$",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "piecewise-linear dynamical systems",
+      "label": "ComputationalModel",
+      "sentence": "Using the fact that the vector fields of the RNN and latent circuit are piecewise-linear dynamical systems, we derive a relationship between their recurrent and input connectivity matrices (Methods):$$Q^{T}{W}_{\\text{rec}}Q={w}_{\\text{rec}}, ,Q^{T}{W}_{\\text{in}}={w}_{\\text{in}}.$$",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "The relation Eq. (5) shows that the latent circuit connectivity w_{rec} is a low-rank structure in the connectivity of a high-dimensional network, which captures interactions among the latent variables defined by the columns of Q.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "The relation between connectivity matrices has the powerful consequence that we can validate the latent circuit mechanism directly in the RNN connectivity.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "First, if the latent circuit faithfully describes the mechanism operating in the RNN, by conjugating the RNN connectivity matrix with Q (Eq. (5)), we expect to find low-dimensional connectivity structure similar to the latent circuit connectivity.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN connectivity matrix",
+      "label": "NeuralCircuit",
+      "sentence": "First, if the latent circuit faithfully describes the mechanism operating in the RNN, by conjugating the RNN connectivity matrix with Q (Eq. (5)), we expect to find low-dimensional connectivity structure similar to the latent circuit connectivity.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "Such an agreement is nontrivial because the latent circuit inference uses only RNN activity without knowledge of the RNN connectivity.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "Such an agreement is nontrivial because the latent circuit inference uses only RNN activity without knowledge of the RNN connectivity.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "Such an agreement is nontrivial because the latent circuit inference uses only RNN activity without knowledge of the RNN connectivity.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "Latent circuit for context-dependent decision-making",
+      "paper_location": "Results"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue^{19} (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue^{19} (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue^{19} (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory stimulus",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue^{19} (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context cue",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue^{19} (Fig. 2a,b and Methods).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "The RNN successfully learns the task; it makes choices according to the relevant stimulus and ignores the irrelevant stimulus in each context (Fig. 2c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "relevant stimulus",
+      "label": "Stimulus",
+      "sentence": "The RNN successfully learns the task; it makes choices according to the relevant stimulus and ignores the irrelevant stimulus in each context (Fig. 2c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant stimulus",
+      "label": "Stimulus",
+      "sentence": "The RNN successfully learns the task; it makes choices according to the relevant stimulus and ignores the irrelevant stimulus in each context (Fig. 2c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalUnit",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "heterogeneous mixed selectivity",
+      "label": "Phenomenon",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "The structure in the RNN connectivity responsible for generating the correct behavioral outputs is not immediately obvious (Fig. 2d).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "behavioral outputs",
+      "label": "Measurement",
+      "sentence": "The structure in the RNN connectivity responsible for generating the correct behavioral outputs is not immediately obvious (Fig. 2d).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We fitted the latent circuit model to the responses of RNN units and its output behavior during the task.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalUnit",
+      "sentence": "We fitted the latent circuit model to the responses of RNN units and its output behavior during the task.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalUnit",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory color nodes",
+      "label": "ComputationalUnit",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory motion nodes",
+      "label": "ComputationalUnit",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "choice nodes",
+      "label": "ComputationalUnit",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "This choice of the latent circuit dimensionality agrees with the dimensionality of RNN responses after training, which is usually close to the total number of inputs and outputs (the first eight principal components accounted for  $97.9\\%$  of the total variance in RNN responses; Supplementary Fig. 2).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "This choice of the latent circuit dimensionality agrees with the dimensionality of RNN responses after training, which is usually close to the total number of inputs and outputs (the first eight principal components accounted for  $97.9\\%$  of the total variance in RNN responses; Supplementary Fig. 2).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "The fitted latent circuit model captured an overwhelming amount of variance in the RNN activity (coefficient of determination  $r^2 = 0.96$  on test data) and accurately matched projected RNN trajectories (Supplementary Fig. 3).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "coefficient of determination",
+      "label": "Measurement",
+      "sentence": "The fitted latent circuit model captured an overwhelming amount of variance in the RNN activity (coefficient of determination  $r^2 = 0.96$  on test data) and accurately matched projected RNN trajectories (Supplementary Fig. 3).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN trajectories",
+      "label": "Measurement",
+      "sentence": "The fitted latent circuit model captured an overwhelming amount of variance in the RNN activity (coefficient of determination  $r^2 = 0.96$  on test data) and accurately matched projected RNN trajectories (Supplementary Fig. 3).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "recurrent connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "The inferred recurrent connectivity  $w_{\\mathrm{rec}}$  of the latent circuit revealed an interpretable mechanism for context-dependent decision-making (Fig. 3a,b).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "The inferred recurrent connectivity  $w_{\\mathrm{rec}}$  of the latent circuit revealed an interpretable mechanism for context-dependent decision-making (Fig. 3a,b).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalUnit",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "left motion",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "green color",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "left choice node",
+      "label": "ComputationalUnit",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "right motion",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "red color",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "right choice node",
+      "label": "ComputationalUnit",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color context node",
+      "label": "ComputationalUnit",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalUnit",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion context node",
+      "label": "ComputationalUnit",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "This pattern of connections from the context nodes to the sensory nodes implements a suppression mechanism that inhibits the irrelevant stimulus-response mapping in each context.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant sensory representation",
+      "label": "Phenomenon",
+      "sentence": "Because the irrelevant sensory representation is suppressed, it does not drive the decision output.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "decision output",
+      "label": "Measurement",
+      "sentence": "Because the irrelevant sensory representation is suppressed, it does not drive the decision output.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "This suppression mechanism based on inhibition of irrelevant representations is qualitatively similar to mechanisms for context-dependent decision-making hypothesized in previous hand-crafted neural circuit models $^{4,9}$ .",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "This suppression mechanism based on inhibition of irrelevant representations is qualitatively similar to mechanisms for context-dependent decision-making hypothesized in previous hand-crafted neural circuit models $^{4,9}$ .",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "This suppression mechanism based on inhibition of irrelevant representations is qualitatively similar to mechanisms for context-dependent decision-making hypothesized in previous hand-crafted neural circuit models $^{4,9}$ .",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "First, we verified the signatures of the suppression mechanism in the RNN activity.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "We projected RNN responses onto the columns of  $Q$ , which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity patterns",
+      "label": "Measurement",
+      "sentence": "We projected RNN responses onto the columns of  $Q$ , which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "We projected RNN responses onto the columns of  $Q$ , which represent RNN activity patterns that correlate with the activity of nodes in the latent circuit (Fig. 3c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion context",
+      "label": "ExperimentalCondition",
+      "sentence": "By projecting RNN responses onto the difference of two columns of  $Q$  corresponding to the context nodes, we obtain a one-dimensional latent variable correlated with the activity difference of the motion context and color context nodes in the latent circuit.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color context nodes",
+      "label": "ComputationalUnit",
+      "sentence": "By projecting RNN responses onto the difference of two columns of  $Q$  corresponding to the context nodes, we obtain a one-dimensional latent variable correlated with the activity difference of the motion context and color context nodes in the latent circuit.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "By projecting RNN responses onto the difference of two columns of  $Q$  corresponding to the context nodes, we obtain a one-dimensional latent variable correlated with the activity difference of the motion context and color context nodes in the latent circuit.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN trajectories",
+      "label": "Measurement",
+      "sentence": "This projection shows RNN trajectories diverging into opposite directions in state space according to context.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "choice axis",
+      "label": "Measurement",
+      "sentence": "Next, the choice axis is the difference of two columns of  $Q$  corresponding to the left and right choice nodes in the latent circuit.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "Projecting RNN activity onto the choice axis reveals trajectories separating according to choice regardless of context.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion axis",
+      "label": "Measurement",
+      "sentence": "Further, the motion axis is the difference of columns of  $Q$  corresponding to the left and right motion nodes, and the color axis is the difference of columns of  $Q$  corresponding to the red and green color nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color axis",
+      "label": "Measurement",
+      "sentence": "Further, the motion axis is the difference of columns of  $Q$  corresponding to the left and right motion nodes, and the color axis is the difference of columns of  $Q$  corresponding to the red and green color nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "Projections of RNN activity",
+      "label": "Measurement",
+      "sentence": "Projections of RNN activity onto the motion and color axes reveal representations of relevant sensory stimuli, whereas representations of irrelevant stimuli are suppressed.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "relevant sensory stimuli",
+      "label": "Stimulus",
+      "sentence": "Projections of RNN activity onto the motion and color axes reveal representations of relevant sensory stimuli, whereas representations of irrelevant stimuli are suppressed.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "Projections of RNN activity onto the motion and color axes reveal representations of relevant sensory stimuli, whereas representations of irrelevant stimuli are suppressed.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color axis",
+      "label": "Measurement",
+      "sentence": "In particular, along the color axis, RNN trajectories separate according to color coherence only on color context trials, whereas on motion context trials, the activity along this axis is suppressed.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN trajectories",
+      "label": "Measurement",
+      "sentence": "In particular, along the color axis, RNN trajectories separate according to color coherence only on color context trials, whereas on motion context trials, the activity along this axis is suppressed.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color coherence",
+      "label": "Stimulus",
+      "sentence": "In particular, along the color axis, RNN trajectories separate according to color coherence only on color context trials, whereas on motion context trials, the activity along this axis is suppressed.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color context trials",
+      "label": "ExperimentalCondition",
+      "sentence": "In particular, along the color axis, RNN trajectories separate according to color coherence only on color context trials, whereas on motion context trials, the activity along this axis is suppressed.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion context trials",
+      "label": "ExperimentalCondition",
+      "sentence": "In particular, along the color axis, RNN trajectories separate according to color coherence only on color context trials, whereas on motion context trials, the activity along this axis is suppressed.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion axis",
+      "label": "Measurement",
+      "sentence": "Similarly, activity along the motion axis is suppressed on color context trials.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color context trials",
+      "label": "ExperimentalCondition",
+      "sentence": "Similarly, activity along the motion axis is suppressed on color context trials.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant sensory representations",
+      "label": "Phenomenon",
+      "sentence": "The suppression of irrelevant sensory representations in RNN activity is consistent with the inhibitory mechanism revealed in the latent circuit connectivity  $w_{\\mathrm{rec}}$ .",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "The suppression of irrelevant sensory representations in RNN activity is consistent with the inhibitory mechanism revealed in the latent circuit connectivity  $w_{\\mathrm{rec}}$ .",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "The suppression of irrelevant sensory representations in RNN activity is consistent with the inhibitory mechanism revealed in the latent circuit connectivity  $w_{\\mathrm{rec}}$ .",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "connectivity relationships",
+      "label": "NeuralCircuit",
+      "sentence": "Second, we used the connectivity relationships Eq. (5) to directly validate the latent circuit mechanism in the RNN connectivity.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "Second, we used the connectivity relationships Eq. (5) to directly validate the latent circuit mechanism in the RNN connectivity.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity matrices",
+      "label": "NeuralCircuit",
+      "sentence": "We conjugated the RNN connectivity matrices with the embedding matrix  $Q$ .",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit",
+      "label": "ComputationalModel",
+      "sentence": "The resulting matrices closely match the connectivity in the latent circuit (Fig. 3d; correlation coefficient  $r = 0.89$ ).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "correlation coefficient",
+      "label": "Measurement",
+      "sentence": "The resulting matrices closely match the connectivity in the latent circuit (Fig. 3d; correlation coefficient  $r = 0.89$ ).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent connectivity structure",
+      "label": "NeuralCircuit",
+      "sentence": "This agreement confirms that the latent connectivity structure indeed exists in the RNN.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "This agreement confirms that the latent connectivity structure indeed exists in the RNN.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "Finally, to ultimately validate that this latent connectivity structure supports the behavioral task performance, we tested whether patterned perturbations of the RNN connectivity (Eq. (6)) produced the same behavioral effects as predicted by the latent circuit model.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "behavioral effects",
+      "label": "Measurement",
+      "sentence": "Finally, to ultimately validate that this latent connectivity structure supports the behavioral task performance, we tested whether patterned perturbations of the RNN connectivity (Eq. (6)) produced the same behavioral effects as predicted by the latent circuit model.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "Finally, to ultimately validate that this latent connectivity structure supports the behavioral task performance, we tested whether patterned perturbations of the RNN connectivity (Eq. (6)) produced the same behavioral effects as predicted by the latent circuit model.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "psychometric function",
+      "label": "Measurement",
+      "sentence": "Indeed, weakening these connections in the latent circuit produced the predicted behavioral effect in the psychometric function, visible as a rotation of the decision boundary on motion context trials (Fig. 4a).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "decision boundary",
+      "label": "Measurement",
+      "sentence": "Indeed, weakening these connections in the latent circuit produced the predicted behavioral effect in the psychometric function, visible as a rotation of the decision boundary on motion context trials (Fig. 4a).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion context trials",
+      "label": "ExperimentalCondition",
+      "sentence": "Indeed, weakening these connections in the latent circuit produced the predicted behavioral effect in the psychometric function, visible as a rotation of the decision boundary on motion context trials (Fig. 4a).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "Perturbations of the RNN connectivity along the corresponding pattern produced similar behavioral effects (Fig. 4b), confirming that this connectivity pattern implements suppression of irrelevant sensory representations in the RNN.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant sensory representations",
+      "label": "Phenomenon",
+      "sentence": "Perturbations of the RNN connectivity along the corresponding pattern produced similar behavioral effects (Fig. 4b), confirming that this connectivity pattern implements suppression of irrelevant sensory representations in the RNN.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Perturbations of the RNN connectivity along the corresponding pattern produced similar behavioral effects (Fig. 4b), confirming that this connectivity pattern implements suppression of irrelevant sensory representations in the RNN.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "red color node",
+      "label": "ComputationalUnit",
+      "sentence": "The latent circuit mechanism predicts that weakening the excitatory connection from the red color node to the right choice node (Fig. 4c) would impair the network's ability to make right choices on the color context trials.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "right choice node",
+      "label": "ComputationalUnit",
+      "sentence": "The latent circuit mechanism predicts that weakening the excitatory connection from the red color node to the right choice node (Fig. 4c) would impair the network's ability to make right choices on the color context trials.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color context trials",
+      "label": "ExperimentalCondition",
+      "sentence": "The latent circuit mechanism predicts that weakening the excitatory connection from the red color node to the right choice node (Fig. 4c) would impair the network's ability to make right choices on the color context trials.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "frequency of right choices",
+      "label": "Measurement",
+      "sentence": "Weakening this connection in the latent circuit indeed decreased the frequency of right choices on color context trials (Fig. 4c).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "This perturbation maps onto a rank-one connectivity perturbation in the RNN, which produced similar behavioral effects (Fig. 4d).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent sensory representation",
+      "label": "Phenomenon",
+      "sentence": "This result confirms both the behavioral relevance of the latent sensory representation and the excitatory mechanism by which it drives choices in the RNN.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "This result confirms both the behavioral relevance of the latent sensory representation and the excitatory mechanism by which it drives choices in the RNN.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Together, these results confirm that the RNN uses the suppression mechanism in which context representations inhibit irrelevant sensory representations.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "Together, these results confirm that the RNN uses the suppression mechanism in which context representations inhibit irrelevant sensory representations.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context representations",
+      "label": "Phenomenon",
+      "sentence": "Together, these results confirm that the RNN uses the suppression mechanism in which context representations inhibit irrelevant sensory representations.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant sensory representations",
+      "label": "Phenomenon",
+      "sentence": "Together, these results confirm that the RNN uses the suppression mechanism in which context representations inhibit irrelevant sensory representations.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We next asked whether RNNs trained on the same task arrive at different circuit solutions for context-dependent decision-making.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "We next asked whether RNNs trained on the same task arrive at different circuit solutions for context-dependent decision-making.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "The latent circuit inference enables us to determine whether different RNNs use the same circuit mechanism.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit inference enables us to determine whether different RNNs use the same circuit mechanism.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "To explore the space of circuit mechanisms, we trained an ensemble of 200 RNNs with randomly initialized connectivity to the same level of task performance  $(r^2 = 0.93 \\pm 0.01$ , coefficient of determination for the RNN and target outputs, mean  $\\pm$  s.d. across networks; Supplementary Fig. 4).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "task performance",
+      "label": "Measurement",
+      "sentence": "To explore the space of circuit mechanisms, we trained an ensemble of 200 RNNs with randomly initialized connectivity to the same level of task performance  $(r^2 = 0.93 \\pm 0.01$ , coefficient of determination for the RNN and target outputs, mean  $\\pm$  s.d. across networks; Supplementary Fig. 4).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "principal component analysis",
+      "label": "Method",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices  $w_{rec}$  and projected the data onto the first two principal components, which accounted for  $42\\%$  of total variance (Fig. 5a).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices  $w_{rec}$  and projected the data onto the first two principal components, which accounted for  $42\\%$  of total variance (Fig. 5a).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "Gaussian mixture model",
+      "label": "Method",
+      "sentence": "The projection reveals three major solution clusters, which we partitioned by fitting data with a Gaussian mixture model with three components.",
+      "paper_location": "Fig. 5 caption"
+    },
+    {
+      "entity": "inhibitory suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses $^{19}$ .",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses $^{19}$ .",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "dynamical selection vector mechanism",
+      "label": "Phenomenon",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses $^{19}$ .",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "Measurement",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses $^{19}$ .",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Representations of irrelevant stimuli in the PFC",
+      "paper_location": "Results"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed $^{19,21,31}$ .",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed $^{19,21,31}$ .",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed $^{19,21,31}$ .",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "Measurement",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for context-dependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed $^{19,21,31}$ .",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "This conclusion was derived using dimensionality reduction methods that fit neural responses with regression models $^{19,21,43-45}$  to find low-dimensional projections that best correlate with task variables (Fig. 1b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "This conclusion was derived using dimensionality reduction methods that fit neural responses with regression models $^{19,21,43-45}$  to find low-dimensional projections that best correlate with task variables (Fig. 1b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "regression models",
+      "label": "Method",
+      "sentence": "This conclusion was derived using dimensionality reduction methods that fit neural responses with regression models $^{19,21,43-45}$  to find low-dimensional projections that best correlate with task variables (Fig. 1b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "By contrast, representations of task variables in the latent circuit model interact via recurrent connectivity to implement the computations necessary to solve the task.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "recurrent connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "By contrast, representations of task variables in the latent circuit model interact via recurrent connectivity to implement the computations necessary to solve the task.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "We therefore sought to determine whether the latent circuit and regression models identify different task representations in the same PFC responses.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC recordings",
+      "label": "Measurement",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies[19,21].",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies[19,21].",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC neurons",
+      "label": "CellType",
+      "sentence": "The dataset consists of several hundred PFC neurons ( $n = 727$  and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task[19] (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "The dataset consists of several hundred PFC neurons ( $n = 727$  and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task[19] (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "rhesus monkeys",
+      "label": "Species",
+      "sentence": "The dataset consists of several hundred PFC neurons ( $n = 727$  and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task[19] (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "The dataset consists of several hundred PFC neurons ( $n = 727$  and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task[19] (Fig. 2a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting  $100\\mathrm{ms}$  after the stimulus onset[19] (Methods).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "750-ms window",
+      "label": "TemporalParameter",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting  $100\\mathrm{ms}$  after the stimulus onset[19] (Methods).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "stimulus onset",
+      "label": "Stimulus",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting  $100\\mathrm{ms}$  after the stimulus onset[19] (Methods).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "The latent circuit model provided good fits of PFC responses projected onto the low-dimensional subspace spanned by the columns of the inferred embedding matrix  $Q$  for both monkeys ( $r^2 = 0.88$  and 0.76 on test data for monkeys A and F, respectively; Extended Data Figs. 7a and 8a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "The latent circuit model provided good fits of PFC responses projected onto the low-dimensional subspace spanned by the columns of the inferred embedding matrix  $Q$  for both monkeys ( $r^2 = 0.88$  and 0.76 on test data for monkeys A and F, respectively; Extended Data Figs. 7a and 8a).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "The task subspace  $Q$  explained a smaller fraction of total variance in the PFC responses ( $11.0\\%$  and  $7.0\\%$  for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC[42] and reflects the high dimensionality of PFC activity[19].",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "The task subspace  $Q$  explained a smaller fraction of total variance in the PFC responses ( $11.0\\%$  and  $7.0\\%$  for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC[42] and reflects the high dimensionality of PFC activity[19].",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "The task subspace  $Q$  explained a smaller fraction of total variance in the PFC responses ( $11.0\\%$  and  $7.0\\%$  for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC[42] and reflects the high dimensionality of PFC activity[19].",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The task subspace  $Q$  explained a smaller fraction of total variance in the PFC responses ( $11.0\\%$  and  $7.0\\%$  for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC[42] and reflects the high dimensionality of PFC activity[19].",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC activity",
+      "label": "Measurement",
+      "sentence": "The task subspace  $Q$  explained a smaller fraction of total variance in the PFC responses ( $11.0\\%$  and  $7.0\\%$  for monkeys A and F, respectively) than in RNNs (Supplementary Fig. 2), which is comparable to previous reports of task-relevant variance in the PFC[42] and reflects the high dimensionality of PFC activity[19].",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann-Whitney  $U$ -test; monkey A: motion  $P = 1.4 \\times 10^{-9}$  and  $U = 120$ , color  $P = 10^{-10}$  and  $U = 83$ ; monkey F: motion  $P = 1.5 \\times 10^{-5}$  and  $U = 277$ , color  $P = 1.6 \\times 10^{-7}$  and  $U = 194$ ;  $n = 36$ ).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann-Whitney  $U$ -test; monkey A: motion  $P = 1.4 \\times 10^{-9}$  and  $U = 120$ , color  $P = 10^{-10}$  and  $U = 83$ ; monkey F: motion  $P = 1.5 \\times 10^{-5}$  and  $U = 277$ , color  $P = 1.6 \\times 10^{-7}$  and  $U = 194$ ;  $n = 36$ ).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "stimulus representations",
+      "label": "Phenomenon",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann-Whitney  $U$ -test; monkey A: motion  $P = 1.4 \\times 10^{-9}$  and  $U = 120$ , color  $P = 10^{-10}$  and  $U = 83$ ; monkey F: motion  $P = 1.5 \\times 10^{-5}$  and  $U = 277$ , color  $P = 1.6 \\times 10^{-7}$  and  $U = 194$ ;  $n = 36$ ).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "Mann-Whitney  $U$ -test",
+      "label": "StatisticalTest",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann-Whitney  $U$ -test; monkey A: motion  $P = 1.4 \\times 10^{-9}$  and  $U = 120$ , color  $P = 10^{-10}$  and  $U = 83$ ; monkey F: motion  $P = 1.5 \\times 10^{-5}$  and  $U = 277$ , color  $P = 1.6 \\times 10^{-7}$  and  $U = 194$ ;  $n = 36$ ).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkey A",
+      "label": "AnimalSubject",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann-Whitney  $U$ -test; monkey A: motion  $P = 1.4 \\times 10^{-9}$  and  $U = 120$ , color  $P = 10^{-10}$  and  $U = 83$ ; monkey F: motion  $P = 1.5 \\times 10^{-5}$  and  $U = 277$ , color  $P = 1.6 \\times 10^{-7}$  and  $U = 194$ ;  $n = 36$ ).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkey F",
+      "label": "AnimalSubject",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann-Whitney  $U$ -test; monkey A: motion  $P = 1.4 \\times 10^{-9}$  and  $U = 120$ , color  $P = 10^{-10}$  and  $U = 83$ ; monkey F: motion  $P = 1.5 \\times 10^{-5}$  and  $U = 277$ , color  $P = 1.6 \\times 10^{-7}$  and  $U = 194$ ;  $n = 36$ ).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC activity",
+      "label": "Measurement",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "latent circuit connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "Consistent with the suppression seen in projected PFC activity, the inferred latent circuit connectivity showed inhibitory connections from context nodes to sensory nodes representing irrelevant stimuli in each context (Fig. 6b and Extended Data Fig. 9b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "We confirmed that PFC responses significantly constrained the inferred latent connectivity above the effect of the task (Extended Data Figs. 7c and 8c and Methods), indicating that the suppression mechanism reflects dynamics in the PFC data.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "latent connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "We confirmed that PFC responses significantly constrained the inferred latent connectivity above the effect of the task (Extended Data Figs. 7c and 8c and Methods), indicating that the suppression mechanism reflects dynamics in the PFC data.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "We confirmed that PFC responses significantly constrained the inferred latent connectivity above the effect of the task (Extended Data Figs. 7c and 8c and Methods), indicating that the suppression mechanism reflects dynamics in the PFC data.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC data",
+      "label": "Measurement",
+      "sentence": "We confirmed that PFC responses significantly constrained the inferred latent connectivity above the effect of the task (Extended Data Figs. 7c and 8c and Methods), indicating that the suppression mechanism reflects dynamics in the PFC data.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC activity",
+      "label": "Measurement",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts[19,21].",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts[19,21].",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "regression methods",
+      "label": "Method",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts[19,21].",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC data",
+      "label": "Measurement",
+      "sentence": "The latent circuit model identified a subspace in PFC activity in which representations of irrelevant stimuli were suppressed, whereas regression methods uncovered subspaces in the same PFC data in which stimuli were nearly equally represented across contexts[19,21].",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "error trials",
+      "label": "ExperimentalCondition",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "incongruent trials",
+      "label": "ExperimentalCondition",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "attention lapse",
+      "label": "Phenotype",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "Method",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "Method",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Stimulus",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "decoding weights",
+      "label": "Measurement",
+      "sentence": "The decoding weights provide an axis in the RNN state space such that a projection onto this axis correlates with the motion coherence.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN state space",
+      "label": "Measurement",
+      "sentence": "The decoding weights provide an axis in the RNN state space such that a projection onto this axis correlates with the motion coherence.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Stimulus",
+      "sentence": "The decoding weights provide an axis in the RNN state space such that a projection onto this axis correlates with the motion coherence.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "decoder axis",
+      "label": "Measurement",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "irrelevant motion stimulus",
+      "label": "Stimulus",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "color context trials",
+      "label": "ExperimentalCondition",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Thus, as in the PFC, irrelevant sensory representations in our RNN appear not suppressed along the decoder axis, whereas they appear suppressed along the axis obtained from the latent circuit model (Fig. 7b).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Thus, as in the PFC, irrelevant sensory representations in our RNN appear not suppressed along the decoder axis, whereas they appear suppressed along the axis obtained from the latent circuit model (Fig. 7b).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "irrelevant sensory representations",
+      "label": "Phenomenon",
+      "sentence": "Thus, as in the PFC, irrelevant sensory representations in our RNN appear not suppressed along the decoder axis, whereas they appear suppressed along the axis obtained from the latent circuit model (Fig. 7b).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "decoder axis",
+      "label": "Measurement",
+      "sentence": "Thus, as in the PFC, irrelevant sensory representations in our RNN appear not suppressed along the decoder axis, whereas they appear suppressed along the axis obtained from the latent circuit model (Fig. 7b).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "Thus, as in the PFC, irrelevant sensory representations in our RNN appear not suppressed along the decoder axis, whereas they appear suppressed along the axis obtained from the latent circuit model (Fig. 7b).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalUnit",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "activity patterns",
+      "label": "Measurement",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "Method",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "psychometric functions",
+      "label": "Measurement",
+      "sentence": "If the corresponding activity patterns are behaviorally relevant, we expect the stimulation to have a substantial effect on psychometric functions.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "right motion stimulus",
+      "label": "Stimulus",
+      "sentence": "Specifically, stimulating the representation of the right motion stimulus should increase the proportion of right choices, shifting the decision boundary on motion context trials.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "decision boundary",
+      "label": "Measurement",
+      "sentence": "Specifically, stimulating the representation of the right motion stimulus should increase the proportion of right choices, shifting the decision boundary on motion context trials.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "motion context trials",
+      "label": "ExperimentalCondition",
+      "sentence": "Specifically, stimulating the representation of the right motion stimulus should increase the proportion of right choices, shifting the decision boundary on motion context trials.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "As expected, driving RNN activity along the motion axis of the latent circuit model shifted the decision boundary on motion context trials (Fig. 7d).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "motion axis",
+      "label": "Measurement",
+      "sentence": "As expected, driving RNN activity along the motion axis of the latent circuit model shifted the decision boundary on motion context trials (Fig. 7d).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "As expected, driving RNN activity along the motion axis of the latent circuit model shifted the decision boundary on motion context trials (Fig. 7d).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "psychometric function",
+      "label": "Measurement",
+      "sentence": "By contrast, stimulation of the same magnitude along the decoder axis had little effect on the psychometric function (Fig. 7c).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "decoder axis",
+      "label": "Measurement",
+      "sentence": "By contrast, stimulation of the same magnitude along the decoder axis had little effect on the psychometric function (Fig. 7c).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "decoder axis",
+      "label": "Measurement",
+      "sentence": "The irrelevant stimulus representations exist along the decoder axis but do not drive the behavioral output.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "The irrelevant stimulus representations exist along the decoder axis but do not drive the behavioral output.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "Single neurons",
+      "label": "CellType",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "heterogeneous responses",
+      "label": "Phenomenon",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "cognitive functions",
+      "label": "CognitiveProcess",
+      "sentence": "Single neurons in higher cortical areas show complex heterogeneous responses during cognitive tasks, posing a challenge for identifying mechanisms of cognitive functions.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "single-neuron heterogeneity",
+      "label": "Phenomenon",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "dimensionality reduction",
+      "label": "Method",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "circuit dynamics",
+      "label": "Phenomenon",
+      "sentence": "Our latent circuit model accounts for single-neuron heterogeneity via dimensionality reduction that incorporates low-dimensional circuit dynamics in its latent variables.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "We show that low-dimensional circuit mechanisms can explain task-relevant dynamics in high-dimensional networks and establish feasibility of inferring these mechanisms from neural response data.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Our theory for interpreting the latent circuit connectivity as a low-rank connectivity in RNNs enables causally validating low-dimensional circuit mechanisms via activity and connectivity perturbations in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "The latent circuit inference can be broadly applied to identify circuit mechanisms for different cognitive tasks from neural response data (Supplementary Fig. 7) and opens new possibilities for causally testing these mechanisms in future experiments.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "latent dynamical systems",
+      "label": "ComputationalModel",
+      "sentence": "Although previous studies extensively modeled neural responses using various latent dynamical systems, our work demonstrates feasibility of fitting neural responses with low-dimensional recurrent circuits.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "Although previous studies extensively modeled neural responses using various latent dynamical systems, our work demonstrates feasibility of fitting neural responses with low-dimensional recurrent circuits.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "low-dimensional recurrent circuits",
+      "label": "NeuralCircuit",
+      "sentence": "Although previous studies extensively modeled neural responses using various latent dynamical systems, our work demonstrates feasibility of fitting neural responses with low-dimensional recurrent circuits.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural data",
+      "label": "Measurement",
+      "sentence": "Most methods fitting low-dimensional dynamical systems to neural data do not explicitly model task inputs and behavior and do not ground the inferred dynamics in the underlying network connectivity[32,33,40,52,53].",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "network connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "Most methods fitting low-dimensional dynamical systems to neural data do not explicitly model task inputs and behavior and do not ground the inferred dynamics in the underlying network connectivity[32,33,40,52,53].",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "RNN architectures",
+      "label": "ComputationalModel",
+      "sentence": "On the other hand, high-dimensional RNN architectures serve as intermediate mechanistic models of neural responses, which generate latent dynamics from high-dimensional recurrent connectivity[25-29].",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "On the other hand, high-dimensional RNN architectures serve as intermediate mechanistic models of neural responses, which generate latent dynamics from high-dimensional recurrent connectivity[25-29].",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "recurrent connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "On the other hand, high-dimensional RNN architectures serve as intermediate mechanistic models of neural responses, which generate latent dynamics from high-dimensional recurrent connectivity[25-29].",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "Our results motivate future work incorporating low-dimensional recurrent circuits as latent dynamics generators within these model architectures, which will enable combining their ability of fitting single-trial neural activity[26,29] with the uniqueness and interpretability furnished by the latent circuit model.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "By relating connectivity to neural dynamics and behavior, the latent circuit model extends the causal predictive power of classical neural circuit models to the study of mechanisms of cognitive functions in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural dynamics",
+      "label": "Phenomenon",
+      "sentence": "By relating connectivity to neural dynamics and behavior, the latent circuit model extends the causal predictive power of classical neural circuit models to the study of mechanisms of cognitive functions in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "By relating connectivity to neural dynamics and behavior, the latent circuit model extends the causal predictive power of classical neural circuit models to the study of mechanisms of cognitive functions in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Although dynamical mechanisms have been studied in RNNs by linearizing the RNN flow field around fixed points $^{19,51}$ , these dynamical mechanisms do not specify how a particular fixed-point configuration arises from the RNN connectivity.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "fixed points",
+      "label": "Phenomenon",
+      "sentence": "Although dynamical mechanisms have been studied in RNNs by linearizing the RNN flow field around fixed points $^{19,51}$ , these dynamical mechanisms do not specify how a particular fixed-point configuration arises from the RNN connectivity.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "NeuralCircuit",
+      "sentence": "Although dynamical mechanisms have been studied in RNNs by linearizing the RNN flow field around fixed points $^{19,51}$ , these dynamical mechanisms do not specify how a particular fixed-point configuration arises from the RNN connectivity.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "irrelevant sensory nodes",
+      "label": "ComputationalUnit",
+      "sentence": "We found that RNNs trained on a context-dependent decision-making task use a suppression mechanism in which context nodes inhibit irrelevant sensory nodes.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "The inhibitory suppression mechanism revealed by the latent circuit model is qualitatively similar to previous neural circuit models of how PFC flexibly switches between alternative stimulus-response mappings $^{8-10,44,45}$ .",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The inhibitory suppression mechanism revealed by the latent circuit model is qualitatively similar to previous neural circuit models of how PFC flexibly switches between alternative stimulus-response mappings $^{8-10,44,45}$ .",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "stimulus-response mappings",
+      "label": "BehavioralParadigm",
+      "sentence": "The inhibitory suppression mechanism revealed by the latent circuit model is qualitatively similar to previous neural circuit models of how PFC flexibly switches between alternative stimulus-response mappings $^{8-10,44,45}$ .",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "Thus, RNNs do not find qualitatively distinct solutions to this task, and complex selectivity of single neurons has a simple explanation as a linear mixing of the low-dimensional latent inhibitory circuit mechanism.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We show that representations of irrelevant stimuli also exist in RNNs that provably implement an inhibitory suppression mechanism, but these representations do not causally drive choices.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "cognitive flexibility",
+      "label": "CognitiveProcess",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "irrelevant stimulus representations",
+      "label": "Phenomenon",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "mean squared error loss function",
+      "label": "StatisticalMethod",
+      "sentence": "We fit the latent circuit model Eqs. (1)—(3) to neural response data y by minimizing the mean squared error loss function,$${L}=\\sum_{k}\\sum_{t}\\parallel {y}_{tk}-Q{x}_{tk}\\parallel_{2}+\\parallel {z}_{tk}-{w}_{out}{x}_{tk}{\\parallel}_{2},$$using custom Python code^{59}.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "We fit the latent circuit model Eqs. (1)—(3) to neural response data y by minimizing the mean squared error loss function,$${L}=\\sum_{k}\\sum_{t}\\parallel {y}_{tk}-Q{x}_{tk}\\parallel_{2}+\\parallel {z}_{tk}-{w}_{out}{x}_{tk}{\\parallel}_{2},$$using custom Python code^{59}.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "We fit the latent circuit model Eqs. (1)—(3) to neural response data y by minimizing the mean squared error loss function,$${L}=\\sum_{k}\\sum_{t}\\parallel {y}_{tk}-Q{x}_{tk}\\parallel_{2}+\\parallel {z}_{tk}-{w}_{out}{x}_{tk}{\\parallel}_{2},$$using custom Python code^{59}.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "nonlinear least squares optimization problem",
+      "label": "StatisticalMethod",
+      "sentence": "Because the variable x depends implicitly on the latent circuit parameters w_{rec} and w_{ou}, the minimization of L is a nonlinear least squares optimization problem^{60} in which we simultaneously search for a behaviorally relevant projection of the high-dimensional activity and a low-dimensional neural circuit that generates dynamics in this projection.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "low-dimensional neural circuit",
+      "label": "NeuralCircuit",
+      "sentence": "Because the variable x depends implicitly on the latent circuit parameters w_{rec} and w_{ou}, the minimization of L is a nonlinear least squares optimization problem^{60} in which we simultaneously search for a behaviorally relevant projection of the high-dimensional activity and a low-dimensional neural circuit that generates dynamics in this projection.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "Cayley transform",
+      "label": "MathematicalMethod",
+      "sentence": "To transform it into an unconstrained problem, we use the Cayley transform to parameterize orthonormal matrices by the linear space of skew symmetric matrices^{61},$$Q=(I+A)(I-A)^{-1}{\\pi }_{n},$$where π_{n} represents projection onto the first n columns, and A is skew symmetric.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "We perform this minimization using PyTorch and the Adam optimizer with default values 0.9 and 0.999 for the decay rate of the first and second moment estimates, respectively, a learning rate of 0.02 and a weight decay of 0.001.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "Adam optimizer",
+      "label": "Software",
+      "sentence": "We perform this minimization using PyTorch and the Adam optimizer with default values 0.9 and 0.999 for the decay rate of the first and second moment estimates, respectively, a learning rate of 0.02 and a weight decay of 0.001.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "Seaborn",
+      "label": "Software",
+      "sentence": "We used the Python software package Seaborn for visualizing model parameters and responses after training.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "entity": "permutation test",
+      "label": "StatisticalTest",
+      "sentence": "To determine whether the inferred latent circuit connectivity significantly depends on neural response data beyond the constraints imposed by the task alone, we performed a permutation test (Extended Data Figs. 1, 7c and 8c), which proceeds in three steps.",
+      "paper_location": "Methods: Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "To determine whether the inferred latent circuit connectivity significantly depends on neural response data beyond the constraints imposed by the task alone, we performed a permutation test (Extended Data Figs. 1, 7c and 8c), which proceeds in three steps.",
+      "paper_location": "Methods: Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "Mann--Whitney U-test",
+      "label": "StatisticalTest",
+      "sentence": "Third, we use a Mann--Whitney U-test to determine whether the correlation coefficients are significantly smaller for models fitted to the shuffled responses than original neural responses.",
+      "paper_location": "Methods: Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "Softplus activation function",
+      "label": "ActivationFunction",
+      "sentence": "To test whether our results extend to networks with a different biologically plausible nonlinearity, we trained RNNs that had a Softplus activation function $f(x)=\\frac{g}{a}\\log(1+e^{gx})$ for a range of parameter β and also with varying gain g across units.",
+      "paper_location": "Methods: Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "To test whether our results extend to networks with a different biologically plausible nonlinearity, we trained RNNs that had a Softplus activation function $f(x)=\\frac{g}{a}\\log(1+e^{gx})$ for a range of parameter β and also with varying gain g across units.",
+      "paper_location": "Methods: Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "rectified linear",
+      "label": "ActivationFunction",
+      "sentence": "We fitted responses of these RNNs with our latent circuit model that had a rectified linear (ReLU) activation function and found that this architecture mismatch did not significantly affect the fit quality and the relationship between connectivity (Extended Data Figs. 5 and 6).",
+      "paper_location": "Methods: Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "ReLU",
+      "label": "ActivationFunction",
+      "sentence": "We fitted responses of these RNNs with our latent circuit model that had a rectified linear (ReLU) activation function and found that this architecture mismatch did not significantly affect the fit quality and the relationship between connectivity (Extended Data Figs. 5 and 6).",
+      "paper_location": "Methods: Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We simulate dynamics of time-discretized RNNs using the general framework for modeling cognitive tasks^{22}.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "We simulate dynamics of time-discretized RNNs using the general framework for modeling cognitive tasks^{22}.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We consider RNNs with positive activity and N = 50 recurrent units.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "recurrent units",
+      "label": "ComputationalUnit",
+      "sentence": "We consider RNNs with positive activity and N = 50 recurrent units.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "first-order Euler scheme",
+      "label": "NumericalMethod",
+      "sentence": "We discretize the RNN dynamics Eq. (10) using the first-order Euler scheme with a time step Δt and add a noise term to obtain$$y_{t}=(1-\\alpha)y_{t-1}+\\alpha\\left[W_{\\mathrm{rec}}y_{t-1}+W_{\\mathrm{in}}u_{t}+\\sqrt{\\frac{2}{\\alpha}}\\sigma_{\\mathrm{rec}}\\xi_{t}\\right]_{+}.$$",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "time constant",
+      "label": "TemporalParameter",
+      "sentence": "We set the time constant τ = 200 ms, the discretization time step Δt = 40 ms, and the noise magnitude σ_{rec} = 0.15.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "200 ms",
+      "label": "TemporalParameter",
+      "sentence": "We set the time constant τ = 200 ms, the discretization time step Δt = 40 ms, and the noise magnitude σ_{rec} = 0.15.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "40 ms",
+      "label": "TemporalParameter",
+      "sentence": "We set the time constant τ = 200 ms, the discretization time step Δt = 40 ms, and the noise magnitude σ_{rec} = 0.15.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "Dale's law",
+      "label": "BiologicalPrinciple",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale's law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "excitatory units",
+      "label": "ComputationalUnit",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale's law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "inhibitory units",
+      "label": "ComputationalUnit",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale's law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "The RNN simulation and training were implemented in Python using the software package PyTorch.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "The RNN simulation and training were implemented in Python using the software package PyTorch.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "context cue",
+      "label": "Stimulus",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "color",
+      "label": "Stimulus",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion context",
+      "label": "ExperimentalCondition",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "sensory stimulus",
+      "label": "Stimulus",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion",
+      "label": "Stimulus",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "color features",
+      "label": "Stimulus",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "right motion",
+      "label": "Stimulus",
+      "sentence": "The right motion and red color are associated with the right choice, and the left motion and green color are associated with the left choice.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "red color",
+      "label": "Stimulus",
+      "sentence": "The right motion and red color are associated with the right choice, and the left motion and green color are associated with the left choice.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "left motion",
+      "label": "Stimulus",
+      "sentence": "The right motion and red color are associated with the right choice, and the left motion and green color are associated with the left choice.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "green color",
+      "label": "Stimulus",
+      "sentence": "The right motion and red color are associated with the right choice, and the left motion and green color are associated with the left choice.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Stimulus",
+      "sentence": "We use motion coherence  $m_{s}$  and color coherence  $c_{s}$  ranging from  $-0.2$  to  $0.2$  chosen from the set  $[-0.2, -0.12, -0.04, 0.04, 0.12, 0.2]$ .",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "color coherence",
+      "label": "Stimulus",
+      "sentence": "We use motion coherence  $m_{s}$  and color coherence  $c_{s}$  ranging from  $-0.2$  to  $0.2$  chosen from the set  $[-0.2, -0.12, -0.04, 0.04, 0.12, 0.2]$ .",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "entity": "Adam algorithm",
+      "label": "Software",
+      "sentence": "The training is performed with the Adam algorithm.",
+      "paper_location": "Methods: RNN training"
+    },
+    {
+      "entity": "Dale's law",
+      "label": "BiologicalPrinciple",
+      "sentence": "To implement Dale's law, connections are clipped to 0 after each training step if they change sign.",
+      "paper_location": "Methods: RNN training"
+    },
+    {
+      "entity": "psychometric functions",
+      "label": "Measurement",
+      "sentence": "Psychometric functions were then computed as the percentage of choices to the right for each combination of context, motion coherence and color coherence.",
+      "paper_location": "Methods: RNN training"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Stimulus",
+      "sentence": "Psychometric functions were then computed as the percentage of choices to the right for each combination of context, motion coherence and color coherence.",
+      "paper_location": "Methods: RNN training"
+    },
+    {
+      "entity": "color coherence",
+      "label": "Stimulus",
+      "sentence": "Psychometric functions were then computed as the percentage of choices to the right for each combination of context, motion coherence and color coherence.",
+      "paper_location": "Methods: RNN training"
+    },
+    {
+      "entity": "linear regression model",
+      "label": "StatisticalMethod",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model\n\n$$\nc = \\beta y + b. \\tag {27}\n$$",
+      "paper_location": "Methods: Linear decoding"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Stimulus",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model\n\n$$\nc = \\beta y + b. \\tag {27}\n$$",
+      "paper_location": "Methods: Linear decoding"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model\n\n$$\nc = \\beta y + b. \\tag {27}\n$$",
+      "paper_location": "Methods: Linear decoding"
+    },
+    {
+      "entity": "regression coefficients",
+      "label": "Measurement",
+      "sentence": "After fitting, we used the vector of regression coefficients  $\\beta$  to define the decoder axis on which we project RNN responses.",
+      "paper_location": "Methods: Linear decoding"
+    },
+    {
+      "entity": "decoder axis",
+      "label": "Measurement",
+      "sentence": "After fitting, we used the vector of regression coefficients  $\\beta$  to define the decoder axis on which we project RNN responses.",
+      "paper_location": "Methods: Linear decoding"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task $^{19}$ .",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "frontal eye field",
+      "label": "BrainRegion",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task $^{19}$ .",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task $^{19}$ .",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task $^{19}$ .",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "neural activity recordings",
+      "label": "Measurement",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task $^{19}$ .",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "762 units",
+      "label": "CellType",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "multiunits",
+      "label": "CellType",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "Monkeys",
+      "label": "Species",
+      "sentence": "Monkeys reported their choice with a saccade to one of two targets presented shortly after fixation for the entire trial duration.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "saccade",
+      "label": "BehavioralResponse",
+      "sentence": "Monkeys reported their choice with a saccade to one of two targets presented shortly after fixation for the entire trial duration.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "fixation",
+      "label": "BehavioralAssay",
+      "sentence": "Monkeys reported their choice with a saccade to one of two targets presented shortly after fixation for the entire trial duration.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "motion direction",
+      "label": "Stimulus",
+      "sentence": "The monkeys were rewarded for saccades to the target location corresponding to the motion direction in the motion context and to the target whose color matched the dominant color of the dots in the color context.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "motion context",
+      "label": "ExperimentalCondition",
+      "sentence": "The monkeys were rewarded for saccades to the target location corresponding to the motion direction in the motion context and to the target whose color matched the dominant color of the dots in the color context.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "dominant color of the dots",
+      "label": "Stimulus",
+      "sentence": "The monkeys were rewarded for saccades to the target location corresponding to the motion direction in the motion context and to the target whose color matched the dominant color of the dots in the color context.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "color context",
+      "label": "ExperimentalCondition",
+      "sentence": "The monkeys were rewarded for saccades to the target location corresponding to the motion direction in the motion context and to the target whose color matched the dominant color of the dots in the color context.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "trial-averaged neural responses",
+      "label": "Measurement",
+      "sentence": "We fitted the latent circuit model to trial-averaged neural responses on correct trials.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "correct trials",
+      "label": "ExperimentalCondition",
+      "sentence": "We fitted the latent circuit model to trial-averaged neural responses on correct trials.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "neurons",
+      "label": "CellType",
+      "sentence": "In our analyses, we included neurons that had at least four correct trials for each of the 72 unique trial conditions, which produced 483 neurons for monkey A and 323 neurons for monkey F.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "483 neurons",
+      "label": "CellType",
+      "sentence": "In our analyses, we included neurons that had at least four correct trials for each of the 72 unique trial conditions, which produced 483 neurons for monkey A and 323 neurons for monkey F.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "323 neurons",
+      "label": "CellType",
+      "sentence": "In our analyses, we included neurons that had at least four correct trials for each of the 72 unique trial conditions, which produced 483 neurons for monkey A and 323 neurons for monkey F.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "firing rates",
+      "label": "Measurement",
+      "sentence": "For each trial, we computed time-varying firing rates by counting spikes in a 50-ms sliding square window (50-ms steps).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "spikes",
+      "label": "Measurement",
+      "sentence": "For each trial, we computed time-varying firing rates by counting spikes in a 50-ms sliding square window (50-ms steps).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "50-ms sliding square window",
+      "label": "Method",
+      "sentence": "For each trial, we computed time-varying firing rates by counting spikes in a 50-ms sliding square window (50-ms steps).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "Gaussian kernel",
+      "label": "Method",
+      "sentence": "Within the training and test sets, we z scored and smoothed (Gaussian kernel,  $\\sigma = 40$  ms) the response of each unit.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "Last, to denoise these trajectories, we projected them onto the principal components explaining  $50\\%$  of their total variance (corresponding to the first 40 and 31 principal components for monkeys A and F, respectively).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "PFC data",
+      "label": "Measurement",
+      "sentence": "We fitted latent circuit models to the PFC data following similar procedures as for RNNs.",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "entity": "RNN models",
+      "label": "ComputationalModel",
+      "sentence": "We analyzed data from 200 RNN models trained with random initializations.",
+      "paper_location": "Methods: Statistics and reproducibility"
+    },
+    {
+      "entity": "PFC data",
+      "label": "Measurement",
+      "sentence": "For PFC data, we fitted 200 latent circuit models to neural responses from each monkey.",
+      "paper_location": "Methods: Statistics and reproducibility"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "For PFC data, we fitted 200 latent circuit models to neural responses from each monkey.",
+      "paper_location": "Methods: Statistics and reproducibility"
+    },
+    {
+      "entity": "monkey",
+      "label": "Species",
+      "sentence": "For PFC data, we fitted 200 latent circuit models to neural responses from each monkey.",
+      "paper_location": "Methods: Statistics and reproducibility"
+    },
+    {
+      "entity": "GitHub",
+      "label": "SoftwareRepository",
+      "sentence": "The source code to reproduce the results of this study is publicly available on GitHub (https://github.com/engellab/latentcircuit) and Zenodo (https://zenodo.org/records/14020108) $^{59}$ .",
+      "paper_location": "Code availability"
+    },
+    {
+      "entity": "Zenodo",
+      "label": "SoftwareRepository",
+      "sentence": "The source code to reproduce the results of this study is publicly available on GitHub (https://github.com/engellab/latentcircuit) and Zenodo (https://zenodo.org/records/14020108) $^{59}$ .",
+      "paper_location": "Code availability"
+    },
+    {
+      "entity": "MAP data-acquisition system",
+      "label": "Instrument",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "Plexon Inc.",
+      "label": "Company",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "Spike sorting",
+      "label": "Method",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "principal component analysis",
+      "label": "Method",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "Plexon offline sorter",
+      "label": "Software",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "Plots were generated with Python (version 3.8) and Seaborn (version 0.9.0).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "Seaborn",
+      "label": "Software",
+      "sentence": "Plots were generated with Python (version 3.8) and Seaborn (version 0.9.0).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "Recurrent neural networks were trained using PyTorch (version 2.4).",
+      "paper_location": "Reporting Summary: Software and code"
+    },
+    {
+      "entity": "macaque monkeys",
+      "label": "Species",
+      "sentence": "Findings were successfully replicated in two macaque monkeys.",
+      "paper_location": "Reporting Summary: Life sciences study design"
+    },
+    {
+      "entity": "adult males",
+      "label": "AnimalSubject",
+      "sentence": "The animals were adult males, ages 13 (monkey A) and 8 (monkey F) years old at the time the experiments were conducted.",
+      "paper_location": "Reporting Summary: Animals and other research organisms"
+    },
+    {
+      "entity": "monkey A",
+      "label": "AnimalSubject",
+      "sentence": "The animals were adult males, ages 13 (monkey A) and 8 (monkey F) years old at the time the experiments were conducted.",
+      "paper_location": "Reporting Summary: Animals and other research organisms"
+    },
+    {
+      "entity": "monkey F",
+      "label": "AnimalSubject",
+      "sentence": "The animals were adult males, ages 13 (monkey A) and 8 (monkey F) years old at the time the experiments were conducted.",
+      "paper_location": "Reporting Summary: Animals and other research organisms"
+    },
+    {
+      "entity": "National Institutes of Health",
+      "label": "Organization",
+      "sentence": "All surgical and behavioral procedures conformed to the guidelines established by the National Institutes of Health and were approved by the Institutional Animal Care and Use Committee of Stanford University.",
+      "paper_location": "Reporting Summary: Animals and other research organisms"
+    },
+    {
+      "entity": "Institutional Animal Care and Use Committee of Stanford University",
+      "label": "EthicsCommittee",
+      "sentence": "All surgical and behavioral procedures conformed to the guidelines established by the National Institutes of Health and were approved by the Institutional Animal Care and Use Committee of Stanford University.",
+      "paper_location": "Reporting Summary: Animals and other research organisms"
+    }
+  ],
+  "key_terms": [
+    {
+      "term": "latent circuit model",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "term": "context-dependent decision-making task",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "term": "patterned connectivity perturbations",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "term": "causal interactions among task variables",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Introductory summary"
+    },
+    {
+      "term": "excitation--inhibition balance",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation--inhibition balance^{12}) and thus can be causally validated in experiments^{13--15}.",
+      "paper_location": "Abstract"
+    },
+    {
+      "term": "mixed selectivity",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit's activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "term": "low-rank structure",
+      "sentence": "The relation Eq. (5) shows that the latent circuit connectivity w_{rec} is a low-rank structure in the connectivity of a high-dimensional network, which captures interactions among the latent variables defined by the columns of Q.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "term": "invariant subspace",
+      "sentence": "Rather, it is a weaker condition that the linear subspace defined by Q is an invariant subspace of the high-dimensional recurrent connectivity matrix.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "term": "rank-one perturbation",
+      "sentence": "Specifically, a change in the connection δ_{ij} between nodes i and j in the latent circuit maps onto a rank-one perturbation of the RNN connectivity matrix (Fig. 1e and Methods),$$\\delta_{ij}\\to q_{i}{q}_{j}^{T},$$where q_{i} is the ith column of Q.",
+      "paper_location": "Results: Interpreting latent connectivity"
+    },
+    {
+      "term": "stimulus-response mappings",
+      "sentence": "This pattern of connections from sensory to choice nodes implements two alternative stimulus-response mappings in the task.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "term": "suppression mechanism",
+      "sentence": "This pattern of connections from the context nodes to the sensory nodes implements a suppression mechanism that inhibits the irrelevant stimulus-response mapping in each context.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "term": "inhibition of irrelevant representations",
+      "sentence": "This suppression mechanism based on inhibition of irrelevant representations is qualitatively similar to mechanisms for context-dependent decision-making hypothesized in previous hand-crafted neural circuit models $^{4,9}$ .",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "term": "motion axis",
+      "sentence": "Further, the motion axis is the difference of columns of  $Q$  corresponding to the left and right motion nodes, and the color axis is the difference of columns of  $Q$  corresponding to the red and green color nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "term": "color axis",
+      "sentence": "Further, the motion axis is the difference of columns of  $Q$  corresponding to the left and right motion nodes, and the color axis is the difference of columns of  $Q$  corresponding to the red and green color nodes.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "term": "choice axis",
+      "sentence": "Next, the choice axis is the difference of two columns of  $Q$  corresponding to the left and right choice nodes in the latent circuit.",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "term": "psychometric functions",
+      "sentence": "Psychometric functions show that the RNN successfully learns the task; it responds to relevant stimuli and ignores irrelevant stimuli in each context.",
+      "paper_location": "Fig. 2 caption"
+    },
+    {
+      "term": "decision boundary",
+      "sentence": "Indeed, weakening these connections in the latent circuit produced the predicted behavioral effect in the psychometric function, visible as a rotation of the decision boundary on motion context trials (Fig. 4a).",
+      "paper_location": "Results: Latent circuit for context-dependent decision-making"
+    },
+    {
+      "term": "space of latent circuit solutions",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices  $w_{rec}$  and projected the data onto the first two principal components, which accounted for  $42\\%$  of total variance (Fig. 5a).",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "term": "solution clusters",
+      "sentence": "The projection reveals three major solution clusters, which we partitioned by fitting data with a Gaussian mixture model with three components.",
+      "paper_location": "Fig. 5 caption"
+    },
+    {
+      "term": "asymmetric inhibitory mechanism",
+      "sentence": "RNNs in cluster 3 operate via a similar asymmetric inhibitory mechanism with the flipped biases for the left and right stimulus representations.",
+      "paper_location": "Extended Data Fig. 3 caption"
+    },
+    {
+      "term": "dynamical selection vector mechanism",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses $^{19}$ .",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "term": "local linear description",
+      "sentence": "We analyzed dynamics in our RNNs and found the same selection vector mechanism (Extended Data Fig. 4), indicating that the dynamical selection vector mechanism is a local linear description of the inhibitory circuit mechanism.",
+      "paper_location": "Results: Space of latent circuit mechanisms"
+    },
+    {
+      "term": "regression models",
+      "sentence": "This conclusion was derived using dimensionality reduction methods that fit neural responses with regression models $^{19,21,43-45}$  to find low-dimensional projections that best correlate with task variables (Fig. 1b).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "term": "smoothed condition-averaged PFC responses",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting  $100\\mathrm{ms}$  after the stimulus onset[19] (Methods).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "term": "neural activity on error trials",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "term": "incongruent trials",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results: Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "term": "linear decoder",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "term": "decoder axis",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "term": "demixing representations of task variables",
+      "sentence": "Our results indicate that 'demixing' representations of task variables $^{19,21,42}$  may not be the right objective for identifying behaviorally relevant patterns in neural activity and may provide a misleading picture of computation.",
+      "paper_location": "Results: Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "term": "single-trial neural activity",
+      "sentence": "Our results motivate future work incorporating low-dimensional recurrent circuits as latent dynamics generators within these model architectures, which will enable combining their ability of fitting single-trial neural activity[26,29] with the uniqueness and interpretability furnished by the latent circuit model.",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "low-rank connectivity",
+      "sentence": "This perspective is qualitatively similar to previous work on engineering low-dimensional task solutions in RNNs with low-rank connectivity $^{47,49,51}$ .",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "cognitive flexibility",
+      "sentence": "Thus, inhibitory mechanisms for cognitive flexibility are compatible with the existence of irrelevant stimulus representations in the PFC.",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "circuit dissection methods",
+      "sentence": "Our work translates such circuit dissection methods $^{16,57}$  to high-dimensional networks in which, unlike in small circuits, individual connections do not carry specific functions, but instead these functions arise from distributed connectivity patterns.",
+      "paper_location": "Discussion"
+    },
+    {
+      "term": "mean squared error loss function",
+      "sentence": "We fit the latent circuit model Eqs. (1)—(3) to neural response data y by minimizing the mean squared error loss function,$${L}=\\sum_{k}\\sum_{t}\\parallel {y}_{tk}-Q{x}_{tk}\\parallel_{2}+\\parallel {z}_{tk}-{w}_{out}{x}_{tk}{\\parallel}_{2},$$using custom Python code^{59}.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "term": "Cayley transform",
+      "sentence": "To transform it into an unconstrained problem, we use the Cayley transform to parameterize orthonormal matrices by the linear space of skew symmetric matrices^{61},$$Q=(I+A)(I-A)^{-1}{\\pi }_{n},$$where π_{n} represents projection onto the first n columns, and A is skew symmetric.",
+      "paper_location": "Methods: Fitting a latent circuit model"
+    },
+    {
+      "term": "permutation test",
+      "sentence": "To determine whether the inferred latent circuit connectivity significantly depends on neural response data beyond the constraints imposed by the task alone, we performed a permutation test (Extended Data Figs. 1, 7c and 8c), which proceeds in three steps.",
+      "paper_location": "Methods: Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "term": "time-discretized RNNs",
+      "sentence": "We simulate dynamics of time-discretized RNNs using the general framework for modeling cognitive tasks^{22}.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "term": "Dale's law",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale's law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods: RNN simulations"
+    },
+    {
+      "term": "context cue",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "term": "motion coherence",
+      "sentence": "We use motion coherence  $m_{s}$  and color coherence  $c_{s}$  ranging from  $-0.2$  to  $0.2$  chosen from the set  $[-0.2, -0.12, -0.04, 0.04, 0.12, 0.2]$ .",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "term": "color coherence",
+      "sentence": "We use motion coherence  $m_{s}$  and color coherence  $c_{s}$  ranging from  $-0.2$  to  $0.2$  chosen from the set  $[-0.2, -0.12, -0.04, 0.04, 0.12, 0.2]$ .",
+      "paper_location": "Methods: Context-dependent decision-making task"
+    },
+    {
+      "term": "50-ms sliding square window",
+      "sentence": "For each trial, we computed time-varying firing rates by counting spikes in a 50-ms sliding square window (50-ms steps).",
+      "paper_location": "Methods: Analysis of PFC data"
+    },
+    {
+      "term": "Gaussian kernel",
+      "sentence": "Within the training and test sets, we z scored and smoothed (Gaussian kernel,  $\\sigma = 40$  ms) the response of each unit.",
+      "paper_location": "Methods: Analysis of PFC data"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_openai-gpt-5.5_20260702_185514.json b/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_openai-gpt-5.5_20260702_185514.json
new file mode 100644
index 0000000..3c19e35
--- /dev/null
+++ b/evaluation/ner/direct_llm_call/output/latent-circuit-inference/neuroscience_ner_openai-gpt-5.5_20260702_185514.json
@@ -0,0 +1,2364 @@
+{
+  "source_metadata": {
+    "source_path": "evaluation/ner/latent-circuit-inference/s41593-025-01869-7.pdf",
+    "statistics": {
+      "model": "openai/gpt-5.5",
+      "prompt_file": "/Users/pujatrivedi/Desktop/MIT/structsense/evaluation/ner/direct_llm_call/prompts/extractor_neuroscience_ner.txt",
+      "input_mode": "grobid_tei",
+      "pdf_engine": null,
+      "grobid_url": "http://localhost:8070",
+      "temperature": null,
+      "seed": null,
+      "finish_reason": "stop",
+      "extracted_at": "20260702_185514",
+      "total_entities": 354,
+      "entities_by_label": {
+        "Measurement": 118,
+        "Stimulus": 36,
+        "ComputationalModel": 33,
+        "Method": 28,
+        "BehavioralAssay": 28,
+        "BrainRegion": 17,
+        "ComputationalModelComponent": 16,
+        "Species": 15,
+        "Phenomenon": 14,
+        "Software": 14,
+        "CellType": 12,
+        "NeuralCircuit": 7,
+        "ComputationalMethod": 5,
+        "Synapse": 2,
+        "SoftwareRepository": 2,
+        "BehavioralParadigm": 1,
+        "AnatomicalSpace": 1,
+        "Phenotype": 1,
+        "Behavior": 1,
+        "Dataset": 1,
+        "Instrument": 1,
+        "Organization": 1
+      }
+    }
+  },
+  "entities": [
+    {
+      "entity": "Higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "sensory, cognitive and motor signals",
+      "label": "Measurement",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "task variables",
+      "label": "Measurement",
+      "sentence": "Higher cortical areas carry a wide range of sensory, cognitive and motor signals mixed in heterogeneous responses of single neurons tuned to multiple task variables.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "Dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "task variables",
+      "label": "Measurement",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "connectivity",
+      "label": "Measurement",
+      "sentence": "Dimensionality reduction methods that rely on correlations between neural activity and task variables leave unknown how heterogeneous responses arise from connectivity to drive behavior.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "dimensionality reduction approach",
+      "label": "Method",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "task variables",
+      "label": "Measurement",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "low-dimensional recurrent connectivity",
+      "label": "Measurement",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "recurrent neural networks",
+      "label": "ComputationalModel",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "contextual representations",
+      "label": "Measurement",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "Measurement",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "behavioral effects",
+      "label": "Measurement",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "patterned connectivity perturbations",
+      "label": "Method",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "Measurement",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "We find similar suppression of irrelevant sensory responses in the prefrontal cortex of monkeys performing the same task.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "task variables",
+      "label": "Measurement",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "higher cortical areas",
+      "label": "BrainRegion",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "sensory and contextual signals",
+      "label": "Measurement",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "behavioral response",
+      "label": "Measurement",
+      "sentence": "Cognitive functions depend on higher cortical areas, which integrate diverse sensory and contextual signals to produce a coherent behavioral response.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "excitatory",
+      "label": "CellType",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "inhibitory neurons",
+      "label": "CellType",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "cortical circuits",
+      "label": "NeuralCircuit",
+      "sentence": "These computations result from interactions between excitatory and inhibitory neurons in cortical circuits.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output [1][2][3][4][5][6][7][8][9][10][11] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "excitation",
+      "label": "Phenomenon",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output [1][2][3][4][5][6][7][8][9][10][11] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "inhibition",
+      "label": "Phenomenon",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output [1][2][3][4][5][6][7][8][9][10][11] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural populations",
+      "label": "CellType",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output [1][2][3][4][5][6][7][8][9][10][11] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "task variables",
+      "label": "Measurement",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output [1][2][3][4][5][6][7][8][9][10][11] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "behavioral output",
+      "label": "Measurement",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output [1][2][3][4][5][6][7][8][9][10][11] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "excitation-inhibition balance",
+      "label": "Phenomenon",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation-inhibition balance 12 ) and thus can be causally validated in experiments [13][14][15] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural dynamics",
+      "label": "Phenomenon",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation-inhibition balance 12 ) and thus can be causally validated in experiments [13][14][15] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "behavioral performance",
+      "label": "Measurement",
+      "sentence": "By linking connectivity to neural dynamics and behavioral output, these models can predict changes in dynamics and behavioral performance under perturbations of the circuit structure (for example, changes in excitation-inhibition balance 12 ) and thus can be causally validated in experiments [13][14][15] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "cortex",
+      "label": "BrainRegion",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of singleneuron responses in the cortex.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "singleneuron responses",
+      "label": "Measurement",
+      "sentence": "However, hand-crafted circuit models come short of capturing the complexity and heterogeneity of singleneuron responses in the cortex.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "Single neurons",
+      "label": "CellType",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables [16][17][18][19][20][21] , which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "prefrontal cortex",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables [16][17][18][19][20][21] , which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables [16][17][18][19][20][21] , which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "multiple task variables",
+      "label": "Measurement",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables [16][17][18][19][20][21] , which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "recurrent neural network",
+      "label": "ComputationalModel",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks 19,[22][23][24] or reproduce neural activity data [25][26][27][28][29] by optimizing recurrent connectivity parameters.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks 19,[22][23][24] or reproduce neural activity data [25][26][27][28][29] by optimizing recurrent connectivity parameters.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks 19,[22][23][24] or reproduce neural activity data [25][26][27][28][29] by optimizing recurrent connectivity parameters.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural activity data",
+      "label": "Measurement",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks 19,[22][23][24] or reproduce neural activity data [25][26][27][28][29] by optimizing recurrent connectivity parameters.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "recurrent connectivity parameters",
+      "label": "Measurement",
+      "sentence": "Similar heterogeneous responses emerge in high-dimensional recurrent neural network (RNN) models trained to perform cognitive tasks 19,[22][23][24] or reproduce neural activity data [25][26][27][28][29] by optimizing recurrent connectivity parameters.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks 30 .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural circuit models",
+      "label": "ComputationalModel",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks 30 .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "high-dimensional activity",
+      "label": "Measurement",
+      "sentence": "Although trained RNNs are a class of neural circuit models, the complexity of their high-dimensional activity and connectivity obscures the interpretation of circuit mechanisms in these networks 30 .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "fixed points",
+      "label": "Phenomenon",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs 19,31 or fitting a dynamical system directly to neural response data 27,[32][33][34][35] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "linearized dynamics",
+      "label": "Phenomenon",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs 19,31 or fitting a dynamical system directly to neural response data 27,[32][33][34][35] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs 19,31 or fitting a dynamical system directly to neural response data 27,[32][33][34][35] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "It has been possible to determine dynamical mechanisms of task computations by either characterizing fixed points and linearized dynamics around them in RNNs 19,31 or fitting a dynamical system directly to neural response data 27,[32][33][34][35] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "flow field",
+      "label": "Phenomenon",
+      "sentence": "These approaches, however, provide no insight into how particular features of the flow field in a dynamical system arise from the network connectivity.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "network connectivity",
+      "label": "Measurement",
+      "sentence": "These approaches, however, provide no insight into how particular features of the flow field in a dynamical system arise from the network connectivity.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Thus, high-dimensional RNNs currently serve as intermediate mechanistic models of heterogeneous neural responses, suggest that RNNs and conceivably the PFC implement qualitatively novel task solutions that do not exist in small circuits and emerge only in high-dimensional networks 19,47,48 .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Thus, high-dimensional RNNs currently serve as intermediate mechanistic models of heterogeneous neural responses, suggest that RNNs and conceivably the PFC implement qualitatively novel task solutions that do not exist in small circuits and emerge only in high-dimensional networks 19,47,48 .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Thus, high-dimensional RNNs currently serve as intermediate mechanistic models of heterogeneous neural responses, suggest that RNNs and conceivably the PFC implement qualitatively novel task solutions that do not exist in small circuits and emerge only in high-dimensional networks 19,47,48 .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "task behavior",
+      "label": "Measurement",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "task variables",
+      "label": "Measurement",
+      "sentence": "To bridge this gap, we develop the latent circuit model, a dimensionality reduction approach that jointly fits neural responses and task behavior and incorporates recurrent interactions among task variables to capture a causal mechanism of task execution.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "low-dimensional latent circuit",
+      "label": "NeuralCircuit",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "low-dimensional circuit mechanism",
+      "label": "NeuralCircuit",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "irrelevant sensory representations",
+      "label": "Measurement",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We validated this mechanism by confirming the behavioral effects of patterned perturbations of the RNN activity and connectivity predicted by the latent circuit model.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "We validated this mechanism by confirming the behavioral effects of patterned perturbations of the RNN activity and connectivity predicted by the latent circuit model.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural recordings",
+      "label": "Method",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods 19,21 .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods 19,21 .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods 19,21 .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods 19,21 .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "Moreover, fitting the latent circuit model to neural recordings from the PFC of monkeys performing the same context-dependent decision-making task revealed a qualitatively similar suppression mechanism, in contrast to previous analyses of the same data with correlation-based methods 19,21 .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural population activity",
+      "label": "Measurement",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "low-dimensional latent variables",
+      "label": "Measurement",
+      "sentence": "Accordingly, dimensionality reduction methods are commonly used to reveal representations of low-dimensional latent variables in neural population activity, which reflect computations emerging at the population level.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "targeted dimensionality reduction methods",
+      "label": "Method",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables 19,21,[41][42][43] (Fig. 1a,b).",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural representations",
+      "label": "Measurement",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables 19,21,[41][42][43] (Fig. 1a,b).",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "neural population activity",
+      "label": "Measurement",
+      "sentence": "Therefore, targeted dimensionality reduction methods directly model neural representations of task inputs and behavioral outputs by seeking low-dimensional projections of neural population activity that correlate with external task variables 19,21,[41][42][43] (Fig. 1a,b).",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC's role in context-dependent decision-making.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "The glaring gap between circuit mechanisms and correlation-based dimensionality reduction methods is apparent in studies of the PFC's role in context-dependent decision-making.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "Context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus-response mappings.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "alternative stimulus-response mappings",
+      "label": "BehavioralParadigm",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus-response mappings.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "sensory representations",
+      "label": "Measurement",
+      "sentence": "Most circuit models hypothesize a relatively simple mechanism based on inhibition of sensory representations irrelevant in the current context [8][9][10][44][45][46] .",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "PFC data",
+      "label": "Measurement",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses 19,21 , seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses 19,21 , seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "Measurement",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses 19,21 , seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Body"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "single-neuron heterogeneity",
+      "label": "Measurement",
+      "sentence": "To bridge the gap between dimensionality reduction, circuit mechanisms and single-neuron heterogeneity, we develop a latent circuit model (Fig. 1c).",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "high-dimensional neural responses",
+      "label": "Measurement",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝ N (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝ n as",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "neurons",
+      "label": "CellType",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝ N (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝ n as",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "cognitive task",
+      "label": "BehavioralAssay",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝ N (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝ n as",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "low-dimensional latent variables",
+      "label": "Measurement",
+      "sentence": "Similar to other dimensionality reduction methods, we model high-dimensional neural responses y ∈ ℝ N (N is the number of neurons) during a cognitive task using low-dimensional latent variables x ∈ ℝ n as",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "neural circuit",
+      "label": "NeuralCircuit",
+      "sentence": "The latent variables x are constrained to be nodes in a neural circuit with dynamics",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "rectified linear (ReLU) activation function",
+      "label": "ComputationalMethod",
+      "sentence": "where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "ReLU",
+      "label": "ComputationalMethod",
+      "sentence": "where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "recurrent connectivity",
+      "label": "Measurement",
+      "sentence": "The latent nodes interact via the recurrent connectivity w rec and receive external task inputs u through the input connectivity w in .",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "external task inputs",
+      "label": "Stimulus",
+      "sentence": "The latent nodes interact via the recurrent connectivity w rec and receive external task inputs u through the input connectivity w in .",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "task outputs",
+      "label": "Measurement",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity w out :",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "circuit activity",
+      "label": "Measurement",
+      "sentence": "We also require the latent circuit to perform the task, that is, we can read out the task outputs z from circuit activity via the output connectivity w out :",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. ( 2).",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "task-related neural activity",
+      "label": "Measurement",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. ( 2).",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "low-dimensional subspace",
+      "label": "AnatomicalSpace",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. ( 2).",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "neural circuit",
+      "label": "NeuralCircuit",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. ( 2).",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "latent circuit parameters",
+      "label": "Measurement",
+      "sentence": "We infer the latent circuit parameters (Q, w rec , w in and w out ) from neural activity y by minimizing the loss function L = ∑ k,t ∥y -Qx∥ 2 + ∥z -w out x∥ 2 , where k and t index trials and time within a trial, respectively (Methods).",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "We infer the latent circuit parameters (Q, w rec , w in and w out ) from neural activity y by minimizing the loss function L = ∑ k,t ∥y -Qx∥ 2 + ∥z -w out x∥ 2 , where k and t index trials and time within a trial, respectively (Methods).",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "loss function",
+      "label": "Measurement",
+      "sentence": "We infer the latent circuit parameters (Q, w rec , w in and w out ) from neural activity y by minimizing the loss function L = ∑ k,t ∥y -Qx∥ 2 + ∥z -w out x∥ 2 , where k and t index trials and time within a trial, respectively (Methods).",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via w in , recurrent interactions among latent nodes via w rec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "task inputs",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via w in , recurrent interactions among latent nodes via w rec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "In the latent circuit model, the heterogeneity of single-neuron responses has three possible sources: mixing of task inputs to the latent circuit via w in , recurrent interactions among latent nodes via w rec and linear mixing of representations in single neurons via the embedding Q.",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "cognitive task",
+      "label": "BehavioralAssay",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit inference",
+      "label": "Method",
+      "sentence": "In this context, RNNs optimized to perform a cognitive task provide an ideal setting for testing and validating the latent circuit inference.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "cortex",
+      "label": "BrainRegion",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit's activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit's activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "mixed selectivity",
+      "label": "Phenomenon",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit's activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit's activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "RNNs mimic the heterogeneity and mixed selectivity of neural responses in the cortex during cognitive tasks while providing full access to each unit's activity, network connectivity and behavioral outcomes.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We consider RNNs with dynamics",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "The relation Eq. ( 5) shows that the latent circuit connectivity w rec is a low-rank structure in the connectivity of a high-dimensional network, which captures interactions among the latent variables defined by the columns of Q.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "latent circuit connectivity",
+      "label": "Measurement",
+      "sentence": "The relation Eq. ( 5) shows that the latent circuit connectivity w rec is a low-rank structure in the connectivity of a high-dimensional network, which captures interactions among the latent variables defined by the columns of Q.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "low-rank structure",
+      "label": "Measurement",
+      "sentence": "The relation Eq. ( 5) shows that the latent circuit connectivity w rec is a low-rank structure in the connectivity of a high-dimensional network, which captures interactions among the latent variables defined by the columns of Q.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "Measurement",
+      "sentence": "The relation between connectivity matrices has the powerful consequence that we can validate the latent circuit mechanism directly in the RNN connectivity.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "Such an agreement is nontrivial because the latent circuit inference uses only RNN activity without knowledge of the RNN connectivity.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "Measurement",
+      "sentence": "Such an agreement is nontrivial because the latent circuit inference uses only RNN activity without knowledge of the RNN connectivity.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue 19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue 19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue 19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion feature",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue 19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory stimulus",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue 19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context cue",
+      "label": "Stimulus",
+      "sentence": "We applied our latent circuit inference to RNNs optimized to perform a context-dependent decision-making task, which requires the discrimination of either the color or motion feature of a sensory stimulus depending on the context cue 19 (Fig. 2a,b and Methods).",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalModelComponent",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "mixed selectivity",
+      "label": "Phenomenon",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "task variables",
+      "label": "Measurement",
+      "sentence": "After training, RNN units show heterogeneous mixed selectivity for multiple task variables (Supplementary Fig. 1).",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit model",
+      "label": "ComputationalModel",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "task variables",
+      "label": "Measurement",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "context nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory color nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory motion nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "choice nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "The latent circuit model consisted of eight nodes corresponding to task variables: two context nodes, two sensory color nodes, two sensory motion nodes and two choice nodes.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "This choice of the latent circuit dimensionality agrees with the dimensionality of RNN responses after training, which is usually close to the total number of inputs and outputs (the first eight principal components accounted for 97.9% of the total variance in RNN responses; Supplementary Fig. 2).",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "This choice of the latent circuit dimensionality agrees with the dimensionality of RNN responses after training, which is usually close to the total number of inputs and outputs (the first eight principal components accounted for 97.9% of the total variance in RNN responses; Supplementary Fig. 2).",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "left motion",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "green color",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "right motion",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "red color",
+      "label": "Stimulus",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "excitatory connections",
+      "label": "Synapse",
+      "sentence": "In the latent circuit, sensory nodes representing stimuli associated with the left choice (left motion and green color) have excitatory connections to the left choice node, and sensory nodes representing stimuli associated with the right choice (right motion and red color) have excitatory connections to the right choice node.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color context node",
+      "label": "ComputationalModelComponent",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "inhibitory connections",
+      "label": "Synapse",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "sensory nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion context node",
+      "label": "ComputationalModelComponent",
+      "sentence": "Further, the color context node has inhibitory connections to the sensory nodes representing motion, and the motion context node has inhibitory connections to sensory nodes representing color.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "This pattern of connections from the context nodes to the sensory nodes implements a suppression mechanism that inhibits the irrelevant stimulus-response mapping in each context.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant sensory representation",
+      "label": "Measurement",
+      "sentence": "Because the irrelevant sensory representation is suppressed, it does not drive the decision output.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "First, we verified the signatures of the suppression mechanism in the RNN activity.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "suppression mechanism",
+      "label": "Phenomenon",
+      "sentence": "First, we verified the signatures of the suppression mechanism in the RNN activity.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "choice axis",
+      "label": "Measurement",
+      "sentence": "Next, the choice axis is the difference of two columns of Q corresponding to the left and right choice nodes in the latent circuit.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion axis",
+      "label": "Measurement",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color axis",
+      "label": "Measurement",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color nodes",
+      "label": "ComputationalModelComponent",
+      "sentence": "Further, the motion axis is the difference of columns of Q corresponding to the left and right motion nodes, and the color axis is the difference of columns of Q corresponding to the red and green color nodes.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN trajectories",
+      "label": "Measurement",
+      "sentence": "In particular, along the color axis, RNN trajectories separate according to color coherence only on color context trials, whereas on motion context trials, the activity along this axis is suppressed.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color coherence",
+      "label": "Measurement",
+      "sentence": "In particular, along the color axis, RNN trajectories separate according to color coherence only on color context trials, whereas on motion context trials, the activity along this axis is suppressed.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color context trials",
+      "label": "BehavioralAssay",
+      "sentence": "In particular, along the color axis, RNN trajectories separate according to color coherence only on color context trials, whereas on motion context trials, the activity along this axis is suppressed.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion context trials",
+      "label": "BehavioralAssay",
+      "sentence": "In particular, along the color axis, RNN trajectories separate according to color coherence only on color context trials, whereas on motion context trials, the activity along this axis is suppressed.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion axis",
+      "label": "Measurement",
+      "sentence": "Similarly, activity along the motion axis is suppressed on color context trials.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "color context trials",
+      "label": "BehavioralAssay",
+      "sentence": "Similarly, activity along the motion axis is suppressed on color context trials.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "irrelevant sensory representations",
+      "label": "Measurement",
+      "sentence": "The suppression of irrelevant sensory representations in RNN activity is consistent with the inhibitory mechanism revealed in the latent circuit connectivity w rec .",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "The suppression of irrelevant sensory representations in RNN activity is consistent with the inhibitory mechanism revealed in the latent circuit connectivity w rec .",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "latent circuit connectivity",
+      "label": "Measurement",
+      "sentence": "The suppression of irrelevant sensory representations in RNN activity is consistent with the inhibitory mechanism revealed in the latent circuit connectivity w rec .",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity",
+      "label": "Measurement",
+      "sentence": "Second, we used the connectivity relationships Eq. ( 5) to directly validate the latent circuit mechanism in the RNN connectivity.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "embedding matrix Q",
+      "label": "Measurement",
+      "sentence": "We conjugated the RNN connectivity matrices with the embedding matrix Q.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNN connectivity matrices",
+      "label": "Measurement",
+      "sentence": "We conjugated the RNN connectivity matrices with the embedding matrix Q.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "psychometric function",
+      "label": "Measurement",
+      "sentence": "Indeed, weakening these connections in the latent circuit produced the predicted behavioral effect in the psychometric function, visible as a rotation of the decision boundary on motion context trials (Fig. 4a).",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "motion context trials",
+      "label": "BehavioralAssay",
+      "sentence": "Indeed, weakening these connections in the latent circuit produced the predicted behavioral effect in the psychometric function, visible as a rotation of the decision boundary on motion context trials (Fig. 4a).",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We next asked whether RNNs trained on the same task arrive at different circuit solutions for context-dependent decision-making.",
+      "paper_location": "Results - Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "We next asked whether RNNs trained on the same task arrive at different circuit solutions for context-dependent decision-making.",
+      "paper_location": "Results - Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "principal component analysis",
+      "label": "Method",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices w rec and projected the data onto the first two principal components, which accounted for 42% of total variance (Fig. 5a).",
+      "paper_location": "Results - Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices w rec and projected the data onto the first two principal components, which accounted for 42% of total variance (Fig. 5a).",
+      "paper_location": "Results - Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "Gaussian mixture model",
+      "label": "Method",
+      "sentence": "The projection reveals three major solution clusters, which we partitioned by fitting data with a Gaussian mixture model with three components.",
+      "paper_location": "Figure 5"
+    },
+    {
+      "entity": "Mann-Whitney U-test",
+      "label": "Method",
+      "sentence": "These latent models produced a significantly worse fit when the latent circuit and the target RNN were sampled from different clusters than from the same cluster (Extended Data Fig. 2m; one-sided Mann-Whitney U-test, U = 52, 593, P < 10 -10 ), confirming differences in circuit mechanisms across clusters.",
+      "paper_location": "Results - Space of latent circuit mechanisms"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for contextdependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed 19,21,51 .",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for contextdependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed 19,21,51 .",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "irrelevant sensory responses",
+      "label": "Measurement",
+      "sentence": "Our finding that RNNs use the inhibitory mechanism for contextdependent decision-making appears in conflict with previous work, which suggested that in both RNNs and the PFC, irrelevant sensory responses are not significantly suppressed 19,21,51 .",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "regression models",
+      "label": "Method",
+      "sentence": "This conclusion was derived using dimensionality reduction methods that fit neural responses with regression models 19,21,[41][42][43] to find low-dimensional projections that best correlate with task variables (Fig. 1b).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "This conclusion was derived using dimensionality reduction methods that fit neural responses with regression models 19,21,[41][42][43] to find low-dimensional projections that best correlate with task variables (Fig. 1b).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC recordings",
+      "label": "Method",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies 19,21 .",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies 19,21 .",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context-dependent decision-making",
+      "label": "BehavioralAssay",
+      "sentence": "We fitted the latent circuit model to the same dataset of PFC recordings during context-dependent decision-making as in the previous studies 19,21 .",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC neurons",
+      "label": "CellType",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task 19 (Fig. 2a).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkeys A and F",
+      "label": "Species",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task 19 (Fig. 2a).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "rhesus monkeys",
+      "label": "Species",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task 19 (Fig. 2a).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "The dataset consists of several hundred PFC neurons (n = 727 and 574 for monkeys A and F, respectively) recorded from two rhesus monkeys performing a context-dependent decision-making task 19 (Fig. 2a).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "smoothed condition-averaged PFC responses",
+      "label": "Measurement",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting 100 ms after the stimulus onset 19 (Methods).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "stimulus onset",
+      "label": "Stimulus",
+      "sentence": "We fitted latent circuit models to smoothed condition-averaged PFC responses during a 750-ms window starting 100 ms after the stimulus onset 19 (Methods).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "The latent circuit model provided good fits of PFC responses projected onto the low-dimensional subspace spanned by the columns of the inferred embedding matrix Q for both monkeys (r 2 = 0.88 and 0.76 on test data for monkeys A and F, respectively; Extended Data Figs. 7a and 8a).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "The latent circuit model provided good fits of PFC responses projected onto the low-dimensional subspace spanned by the columns of the inferred embedding matrix Q for both monkeys (r 2 = 0.88 and 0.76 on test data for monkeys A and F, respectively; Extended Data Figs. 7a and 8a).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkeys A and F",
+      "label": "Species",
+      "sentence": "The latent circuit model provided good fits of PFC responses projected onto the low-dimensional subspace spanned by the columns of the inferred embedding matrix Q for both monkeys (r 2 = 0.88 and 0.76 on test data for monkeys A and F, respectively; Extended Data Figs. 7a and 8a).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann-Whitney U-test; monkey A: motion P = 1.4 × 10 -9 and U = 120, color P = 10 -10 and U = 83; monkey F: motion P = 1.5 × 10 -5 and U = 277, color P = 1.6 × 10 -7 and U = 194; n = 36).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann-Whitney U-test; monkey A: motion P = 1.4 × 10 -9 and U = 120, color P = 10 -10 and U = 83; monkey F: motion P = 1.5 × 10 -5 and U = 277, color P = 1.6 × 10 -7 and U = 194; n = 36).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann-Whitney U-test; monkey A: motion P = 1.4 × 10 -9 and U = 120, color P = 10 -10 and U = 83; monkey F: motion P = 1.5 × 10 -5 and U = 277, color P = 1.6 × 10 -7 and U = 194; n = 36).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "stimulus representations",
+      "label": "Measurement",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann-Whitney U-test; monkey A: motion P = 1.4 × 10 -9 and U = 120, color P = 10 -10 and U = 83; monkey F: motion P = 1.5 × 10 -5 and U = 277, color P = 1.6 × 10 -7 and U = 194; n = 36).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "Mann-Whitney U-test",
+      "label": "Method",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann-Whitney U-test; monkey A: motion P = 1.4 × 10 -9 and U = 120, color P = 10 -10 and U = 83; monkey F: motion P = 1.5 × 10 -5 and U = 277, color P = 1.6 × 10 -7 and U = 194; n = 36).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkey A",
+      "label": "Species",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann-Whitney U-test; monkey A: motion P = 1.4 × 10 -9 and U = 120, color P = 10 -10 and U = 83; monkey F: motion P = 1.5 × 10 -5 and U = 277, color P = 1.6 × 10 -7 and U = 194; n = 36).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "monkey F",
+      "label": "Species",
+      "sentence": "Similar to the RNNs, projecting PFC responses onto the axes identified by the latent circuit model revealed significant suppression of stimulus representations when they were irrelevant (Fig. 6a and Extended Data Figs. 9a and 10; one-sided Mann-Whitney U-test; monkey A: motion P = 1.4 × 10 -9 and U = 120, color P = 10 -10 and U = 83; monkey F: motion P = 1.5 × 10 -5 and U = 277, color P = 1.6 × 10 -7 and U = 194; n = 36).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC activity",
+      "label": "Measurement",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "neural activity",
+      "label": "Measurement",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "error trials",
+      "label": "BehavioralAssay",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "incongruent trials",
+      "label": "BehavioralAssay",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "irrelevant stimuli",
+      "label": "Stimulus",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "attention lapse",
+      "label": "Phenotype",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "Method",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We compared projections of RNN activity onto axes obtained from the latent circuit model and a linear decoder.",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN activity",
+      "label": "Measurement",
+      "sentence": "We trained a linear decoder to predict the signed motion coherence on each trial from RNN activity (Methods).",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "decoder axis",
+      "label": "Measurement",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "irrelevant motion stimulus",
+      "label": "Stimulus",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "color context trials",
+      "label": "BehavioralAssay",
+      "sentence": "By projecting RNN responses onto the decoder axis, we find a strong representation of irrelevant motion stimulus on color context trials without noticeable suppression (Fig. 7a).",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "RNN units",
+      "label": "ComputationalModelComponent",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "linear decoder",
+      "label": "Method",
+      "sentence": "To test this idea, we stimulated RNN units with activity patterns aligned with the axes obtained from the linear decoder and the latent circuit model.",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "psychometric functions",
+      "label": "Measurement",
+      "sentence": "If the corresponding activity patterns are behaviorally relevant, we expect the stimulation to have a substantial effect on psychometric functions.",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "right motion stimulus",
+      "label": "Stimulus",
+      "sentence": "Specifically, stimulating the representation of the right motion stimulus should increase the proportion of right choices, shifting the decision boundary on motion context trials.",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "motion context trials",
+      "label": "BehavioralAssay",
+      "sentence": "Specifically, stimulating the representation of the right motion stimulus should increase the proportion of right choices, shifting the decision boundary on motion context trials.",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "We show that low-dimensional circuit mechanisms can explain task-relevant dynamics in high-dimensional networks and establish feasibility of inferring these mechanisms from neural response data.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "Our theory for interpreting the latent circuit connectivity as a low-rank connectivity in RNNs enables causally validating low-dimensional circuit mechanisms via activity and connectivity perturbations in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "low-rank connectivity",
+      "label": "Measurement",
+      "sentence": "Our theory for interpreting the latent circuit connectivity as a low-rank connectivity in RNNs enables causally validating low-dimensional circuit mechanisms via activity and connectivity perturbations in high-dimensional networks.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "low-dimensional dynamical systems",
+      "label": "ComputationalModel",
+      "sentence": "Although previous studies extensively modeled neural responses using various latent dynamical systems, our work demonstrates feasibility of fitting neural responses with low-dimensional recurrent circuits.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "low-dimensional recurrent circuits",
+      "label": "NeuralCircuit",
+      "sentence": "Although previous studies extensively modeled neural responses using various latent dynamical systems, our work demonstrates feasibility of fitting neural responses with low-dimensional recurrent circuits.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "irrelevant sensory representations",
+      "label": "Measurement",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "The latent circuit model revealed a suppression of irrelevant sensory representations in PFC responses of monkeys performing a context-dependent decision-making task.",
+      "paper_location": "Discussion"
+    },
+    {
+      "entity": "mean squared error loss function",
+      "label": "Measurement",
+      "sentence": "We fit the latent circuit model Eqs. (1)-(3) to neural response data y by minimizing the mean squared error loss function,",
+      "paper_location": "Methods - Fitting a latent circuit model"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "We fit the latent circuit model Eqs. (1)-(3) to neural response data y by minimizing the mean squared error loss function,",
+      "paper_location": "Methods - Fitting a latent circuit model"
+    },
+    {
+      "entity": "custom Python code",
+      "label": "Software",
+      "sentence": "using custom Python code 59 .",
+      "paper_location": "Methods - Fitting a latent circuit model"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "using custom Python code 59 .",
+      "paper_location": "Methods - Fitting a latent circuit model"
+    },
+    {
+      "entity": "nonlinear least squares optimization problem",
+      "label": "Method",
+      "sentence": "Because the variable x depends implicitly on the latent circuit parameters w rec and w in , the minimization of L is a nonlinear least squares optimization problem 60 in which we simultaneously search for a behaviorally relevant projection of the high-dimensional activity and a low-dimensional neural circuit that generates dynamics in this projection.",
+      "paper_location": "Methods - Fitting a latent circuit model"
+    },
+    {
+      "entity": "low-dimensional neural circuit",
+      "label": "NeuralCircuit",
+      "sentence": "Because the variable x depends implicitly on the latent circuit parameters w rec and w in , the minimization of L is a nonlinear least squares optimization problem 60 in which we simultaneously search for a behaviorally relevant projection of the high-dimensional activity and a low-dimensional neural circuit that generates dynamics in this projection.",
+      "paper_location": "Methods - Fitting a latent circuit model"
+    },
+    {
+      "entity": "Cayley transform",
+      "label": "Method",
+      "sentence": "To transform it into an unconstrained problem, we use the Cayley transform to parameterize orthonormal matrices by the linear space of skew symmetric matrices 61 ,",
+      "paper_location": "Methods - Fitting a latent circuit model"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "We perform this minimization using PyTorch and the Adam optimizer with default values 0.9 and 0.999 for the decay rate of the first and second moment estimates, respectively, a learning rate of 0.02 and a weight decay of 0.001.",
+      "paper_location": "Methods - Fitting a latent circuit model"
+    },
+    {
+      "entity": "Adam optimizer",
+      "label": "Software",
+      "sentence": "We perform this minimization using PyTorch and the Adam optimizer with default values 0.9 and 0.999 for the decay rate of the first and second moment estimates, respectively, a learning rate of 0.02 and a weight decay of 0.001.",
+      "paper_location": "Methods - Fitting a latent circuit model"
+    },
+    {
+      "entity": "Seaborn",
+      "label": "Software",
+      "sentence": "We used the Python software package Seaborn for visualizing model parameters and responses after training.",
+      "paper_location": "Methods - Fitting a latent circuit model"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "We used the Python software package Seaborn for visualizing model parameters and responses after training.",
+      "paper_location": "Methods - Fitting a latent circuit model"
+    },
+    {
+      "entity": "permutation test",
+      "label": "Method",
+      "sentence": "To determine whether the inferred latent circuit connectivity significantly depends on neural response data beyond the constraints imposed by the task alone, we performed a permutation test (Extended Data Figs. 1, 7c and 8c), which proceeds in three steps.",
+      "paper_location": "Methods - Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "neural response data",
+      "label": "Measurement",
+      "sentence": "To determine whether the inferred latent circuit connectivity significantly depends on neural response data beyond the constraints imposed by the task alone, we performed a permutation test (Extended Data Figs. 1, 7c and 8c), which proceeds in three steps.",
+      "paper_location": "Methods - Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "Mann-Whitney U-test",
+      "label": "Method",
+      "sentence": "Third, we use a Mann-Whitney U-test to determine whether the correlation coefficients are significantly smaller for models fitted to the shuffled responses than original neural responses.",
+      "paper_location": "Methods - Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "RNN",
+      "label": "ComputationalModel",
+      "sentence": "We used the same test for both RNN (N = 500; Extended Data Fig. 1) and PFC data (N = 1,000; Extended Data Figs. 7c and 8c).",
+      "paper_location": "Methods - Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "PFC data",
+      "label": "Measurement",
+      "sentence": "We used the same test for both RNN (N = 500; Extended Data Fig. 1) and PFC data (N = 1,000; Extended Data Figs. 7c and 8c).",
+      "paper_location": "Methods - Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "entity": "rectified linear (ReLU) activation function",
+      "label": "ComputationalMethod",
+      "sentence": "Here, [⋅] + is a rectified linear (ReLU) activation function, τ is a time constant, and u are external task inputs.",
+      "paper_location": "Methods - Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "ReLU",
+      "label": "ComputationalMethod",
+      "sentence": "Here, [⋅] + is a rectified linear (ReLU) activation function, τ is a time constant, and u are external task inputs.",
+      "paper_location": "Methods - Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "external task inputs",
+      "label": "Stimulus",
+      "sentence": "Here, [⋅] + is a rectified linear (ReLU) activation function, τ is a time constant, and u are external task inputs.",
+      "paper_location": "Methods - Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "Softplus activation function",
+      "label": "ComputationalMethod",
+      "sentence": "To test whether our results extend to networks with a different biologically plausible nonlinearity, we trained RNNs that had a Softplus activation function f(x) = g β log(1 + e βx ) for a range of parameter β and also with varying gain g across units.",
+      "paper_location": "Methods - Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "To test whether our results extend to networks with a different biologically plausible nonlinearity, we trained RNNs that had a Softplus activation function f(x) = g β log(1 + e βx ) for a range of parameter β and also with varying gain g across units.",
+      "paper_location": "Methods - Relationship between connectivity of the RNN and latent circuit"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We simulate dynamics of time-discretized RNNs using the general framework for modeling cognitive tasks 22 .",
+      "paper_location": "Methods - RNN simulations"
+    },
+    {
+      "entity": "cognitive tasks",
+      "label": "BehavioralAssay",
+      "sentence": "We simulate dynamics of time-discretized RNNs using the general framework for modeling cognitive tasks 22 .",
+      "paper_location": "Methods - RNN simulations"
+    },
+    {
+      "entity": "recurrent units",
+      "label": "ComputationalModelComponent",
+      "sentence": "We consider RNNs with positive activity and N = 50 recurrent units.",
+      "paper_location": "Methods - RNN simulations"
+    },
+    {
+      "entity": "Dale's law",
+      "label": "Phenomenon",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale's law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods - RNN simulations"
+    },
+    {
+      "entity": "excitatory units",
+      "label": "ComputationalModelComponent",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale's law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods - RNN simulations"
+    },
+    {
+      "entity": "inhibitory units",
+      "label": "ComputationalModelComponent",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale's law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods - RNN simulations"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "The RNN simulation and training were implemented in Python using the software package PyTorch.",
+      "paper_location": "Methods - RNN simulations"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "The RNN simulation and training were implemented in Python using the software package PyTorch.",
+      "paper_location": "Methods - RNN simulations"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "context cue",
+      "label": "Stimulus",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "color",
+      "label": "Stimulus",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion context",
+      "label": "Stimulus",
+      "sentence": "In the context-dependent decision-making task, at the beginning of each trial, a context cue briefly appears to indicate either the color or motion context for the current trial.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "sensory stimulus",
+      "label": "Stimulus",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion",
+      "label": "Stimulus",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "color features",
+      "label": "Stimulus",
+      "sentence": "After a short delay, a sensory stimulus appears that consists of motion and color features.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "right motion",
+      "label": "Stimulus",
+      "sentence": "The right motion and red color are associated with the right choice, and the left motion and green color are associated with the left choice.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "red color",
+      "label": "Stimulus",
+      "sentence": "The right motion and red color are associated with the right choice, and the left motion and green color are associated with the left choice.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "left motion",
+      "label": "Stimulus",
+      "sentence": "The right motion and red color are associated with the right choice, and the left motion and green color are associated with the left choice.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "green color",
+      "label": "Stimulus",
+      "sentence": "The right motion and red color are associated with the right choice, and the left motion and green color are associated with the left choice.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "color coherence",
+      "label": "Measurement",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion and color stimuli",
+      "label": "Stimulus",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion stimulus",
+      "label": "Stimulus",
+      "sentence": "In the color context, the choice should be made according to the color, ignoring the motion stimulus, and vice versa in the motion context.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion context",
+      "label": "BehavioralAssay",
+      "sentence": "In the color context, the choice should be made according to the color, ignoring the motion stimulus, and vice versa in the motion context.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion context",
+      "label": "Stimulus",
+      "sentence": "To model the context-dependent decision-making task, the network receives six inputs u corresponding to two context cues (u m : motion context; u c : color context) and sensory evidence streams for motion (u m,L : motion left; u m,R : motion right) and color (u c,R : color red; u c,G : color green).",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "color context",
+      "label": "Stimulus",
+      "sentence": "To model the context-dependent decision-making task, the network receives six inputs u corresponding to two context cues (u m : motion context; u c : color context) and sensory evidence streams for motion (u m,L : motion left; u m,R : motion right) and color (u c,R : color red; u c,G : color green).",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion left",
+      "label": "Stimulus",
+      "sentence": "To model the context-dependent decision-making task, the network receives six inputs u corresponding to two context cues (u m : motion context; u c : color context) and sensory evidence streams for motion (u m,L : motion left; u m,R : motion right) and color (u c,R : color red; u c,G : color green).",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "motion right",
+      "label": "Stimulus",
+      "sentence": "To model the context-dependent decision-making task, the network receives six inputs u corresponding to two context cues (u m : motion context; u c : color context) and sensory evidence streams for motion (u m,L : motion left; u m,R : motion right) and color (u c,R : color red; u c,G : color green).",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "color red",
+      "label": "Stimulus",
+      "sentence": "To model the context-dependent decision-making task, the network receives six inputs u corresponding to two context cues (u m : motion context; u c : color context) and sensory evidence streams for motion (u m,L : motion left; u m,R : motion right) and color (u c,R : color red; u c,G : color green).",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "color green",
+      "label": "Stimulus",
+      "sentence": "To model the context-dependent decision-making task, the network receives six inputs u corresponding to two context cues (u m : motion context; u c : color context) and sensory evidence streams for motion (u m,L : motion left; u m,R : motion right) and color (u c,R : color red; u c,G : color green).",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "entity": "firing rates",
+      "label": "Measurement",
+      "sentence": "The first term is the task error, and the second term serves to regularize by penalizing the magnitude of the firing rates.",
+      "paper_location": "Methods - RNN training"
+    },
+    {
+      "entity": "Adam algorithm",
+      "label": "Software",
+      "sentence": "The training is performed with the Adam algorithm.",
+      "paper_location": "Methods - RNN training"
+    },
+    {
+      "entity": "psychometric functions",
+      "label": "Measurement",
+      "sentence": "Psychometric functions were then computed as the percentage of choices to the right for each combination of context, motion coherence and color coherence.",
+      "paper_location": "Methods - RNN training"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "Psychometric functions were then computed as the percentage of choices to the right for each combination of context, motion coherence and color coherence.",
+      "paper_location": "Methods - RNN training"
+    },
+    {
+      "entity": "color coherence",
+      "label": "Measurement",
+      "sentence": "Psychometric functions were then computed as the percentage of choices to the right for each combination of context, motion coherence and color coherence.",
+      "paper_location": "Methods - RNN training"
+    },
+    {
+      "entity": "linear regression model",
+      "label": "Method",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model",
+      "paper_location": "Methods - Linear decoding"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model",
+      "paper_location": "Methods - Linear decoding"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "To decode motion coherence from RNN responses, we fit a linear regression model",
+      "paper_location": "Methods - Linear decoding"
+    },
+    {
+      "entity": "regression coefficients",
+      "label": "Measurement",
+      "sentence": "where β ∈ ℝ 1×N is the vector of regression coefficients, c ∈ ℝ 1×K⋅T is the motion coherence on each trial, b ∈ ℝ is a bias term, and y ∈ ℝ N×K⋅T is the RNN responses at each time step during the stimulus epoch of each trial.",
+      "paper_location": "Methods - Linear decoding"
+    },
+    {
+      "entity": "motion coherence",
+      "label": "Measurement",
+      "sentence": "where β ∈ ℝ 1×N is the vector of regression coefficients, c ∈ ℝ 1×K⋅T is the motion coherence on each trial, b ∈ ℝ is a bias term, and y ∈ ℝ N×K⋅T is the RNN responses at each time step during the stimulus epoch of each trial.",
+      "paper_location": "Methods - Linear decoding"
+    },
+    {
+      "entity": "RNN responses",
+      "label": "Measurement",
+      "sentence": "where β ∈ ℝ 1×N is the vector of regression coefficients, c ∈ ℝ 1×K⋅T is the motion coherence on each trial, b ∈ ℝ is a bias term, and y ∈ ℝ N×K⋅T is the RNN responses at each time step during the stimulus epoch of each trial.",
+      "paper_location": "Methods - Linear decoding"
+    },
+    {
+      "entity": "stimulus epoch",
+      "label": "Stimulus",
+      "sentence": "where β ∈ ℝ 1×N is the vector of regression coefficients, c ∈ ℝ 1×K⋅T is the motion coherence on each trial, b ∈ ℝ is a bias term, and y ∈ ℝ N×K⋅T is the RNN responses at each time step during the stimulus epoch of each trial.",
+      "paper_location": "Methods - Linear decoding"
+    },
+    {
+      "entity": "PFC",
+      "label": "BrainRegion",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task 19 .",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "frontal eye field",
+      "label": "BrainRegion",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task 19 .",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "neural activity recordings",
+      "label": "Method",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task 19 .",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "monkeys",
+      "label": "Species",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task 19 .",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "context-dependent decision-making task",
+      "label": "BehavioralAssay",
+      "sentence": "We analyzed a publicly available dataset of neural activity recordings from the PFC (in and around the frontal eye field) from two monkeys performing a context-dependent decision-making task 19 .",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "762 units",
+      "label": "CellType",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "monkey A",
+      "label": "Species",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "640 units",
+      "label": "CellType",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "monkey F",
+      "label": "Species",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "single neurons",
+      "label": "CellType",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "multiunits",
+      "label": "CellType",
+      "sentence": "This dataset consisted of 762 units from monkey A and 640 units from monkey F (including single neurons and multiunits).",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "motion contexts",
+      "label": "BehavioralAssay",
+      "sentence": "Because stimulus coherence levels varied across monkeys and days, to equate performance in the motion and color contexts, we replaced the coherences on each trial with their average values for each stimulus difficulty (average motion coherences: 0.05, 0.15 and 0.50 in monkey A and 0.07, 0.19 and 0.54 in monkey F; average color coherences: 0.06, 0.18 and 0.50 in monkey A and 0.12, 0.30 and 0.75 in monkey F).",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "color contexts",
+      "label": "BehavioralAssay",
+      "sentence": "Because stimulus coherence levels varied across monkeys and days, to equate performance in the motion and color contexts, we replaced the coherences on each trial with their average values for each stimulus difficulty (average motion coherences: 0.05, 0.15 and 0.50 in monkey A and 0.07, 0.19 and 0.54 in monkey F; average color coherences: 0.06, 0.18 and 0.50 in monkey A and 0.12, 0.30 and 0.75 in monkey F).",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "saccade",
+      "label": "Behavior",
+      "sentence": "Monkeys reported their choice with a saccade to one of two targets presented shortly after fixation for the entire trial duration.",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "Monkeys",
+      "label": "Species",
+      "sentence": "Monkeys reported their choice with a saccade to one of two targets presented shortly after fixation for the entire trial duration.",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "random dots stimulus",
+      "label": "Stimulus",
+      "sentence": "We analyzed neural responses during the presentation of the random dots stimulus because the available data consisted of neural responses starting at 100 ms after stimulus onset for a duration of 750 ms.",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "We analyzed neural responses during the presentation of the random dots stimulus because the available data consisted of neural responses starting at 100 ms after stimulus onset for a duration of 750 ms.",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "stimulus onset",
+      "label": "Stimulus",
+      "sentence": "We analyzed neural responses during the presentation of the random dots stimulus because the available data consisted of neural responses starting at 100 ms after stimulus onset for a duration of 750 ms.",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "firing rates",
+      "label": "Measurement",
+      "sentence": "For each trial, we computed time-varying firing rates by counting spikes in a 50-ms sliding square window (50-ms steps).",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "spikes",
+      "label": "Measurement",
+      "sentence": "For each trial, we computed time-varying firing rates by counting spikes in a 50-ms sliding square window (50-ms steps).",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "Gaussian kernel",
+      "label": "Method",
+      "sentence": "Within the training and test sets, we z scored and smoothed (Gaussian kernel, σ = 40 ms) the response of each unit.",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "single-neuron responses",
+      "label": "Measurement",
+      "sentence": "To construct population responses, we combined the single-neuron responses for each trial condition.",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "population responses",
+      "label": "Measurement",
+      "sentence": "To construct population responses, we combined the single-neuron responses for each trial condition.",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "principal components",
+      "label": "Method",
+      "sentence": "Last, to denoise these trajectories, we projected them onto the principal components explaining 50% of their total variance (corresponding to the first 40 and 31 principal components for monkeys A and F, respectively).",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "PFC data",
+      "label": "Measurement",
+      "sentence": "We fitted latent circuit models to the PFC data following similar procedures as for RNNs.",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "RNNs",
+      "label": "ComputationalModel",
+      "sentence": "We fitted latent circuit models to the PFC data following similar procedures as for RNNs.",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "PFC responses",
+      "label": "Measurement",
+      "sentence": "Because of high dimensionality of PFC responses (40 and 31 principal components are required to account for ~50% of the total variance in PFC activity for monkeys A and F, respectively), we find a notable tradeoff between the task fit and data fit when fitting the low-dimensional latent circuit model to the PFC data.",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "PFC activity",
+      "label": "Measurement",
+      "sentence": "Because of high dimensionality of PFC responses (40 and 31 principal components are required to account for ~50% of the total variance in PFC activity for monkeys A and F, respectively), we find a notable tradeoff between the task fit and data fit when fitting the low-dimensional latent circuit model to the PFC data.",
+      "paper_location": "Methods - Analysis of PFC data"
+    },
+    {
+      "entity": "RNN models",
+      "label": "ComputationalModel",
+      "sentence": "We analyzed data from 200 RNN models trained with random initializations.",
+      "paper_location": "Methods - Statistics and reproducibility"
+    },
+    {
+      "entity": "PFC data",
+      "label": "Measurement",
+      "sentence": "For PFC data, we fitted 200 latent circuit models to neural responses from each monkey.",
+      "paper_location": "Methods - Statistics and reproducibility"
+    },
+    {
+      "entity": "neural responses",
+      "label": "Measurement",
+      "sentence": "For PFC data, we fitted 200 latent circuit models to neural responses from each monkey.",
+      "paper_location": "Methods - Statistics and reproducibility"
+    },
+    {
+      "entity": "monkey",
+      "label": "Species",
+      "sentence": "For PFC data, we fitted 200 latent circuit models to neural responses from each monkey.",
+      "paper_location": "Methods - Statistics and reproducibility"
+    },
+    {
+      "entity": "Neural recording data",
+      "label": "Method",
+      "sentence": "Neural recording data were previously described in Mante et al. 19 ; no randomization or blinding was performed because there was only one experimental group.",
+      "paper_location": "Methods - Statistics and reproducibility"
+    },
+    {
+      "entity": "Mante et al.",
+      "label": "Dataset",
+      "sentence": "Neural recording data were previously described in Mante et al. 19 ; no randomization or blinding was performed because there was only one experimental group.",
+      "paper_location": "Methods - Statistics and reproducibility"
+    },
+    {
+      "entity": "macaque monkeys",
+      "label": "Species",
+      "sentence": "Findings were successfully replicated in two macaque monkeys.",
+      "paper_location": "Reporting summary - Replication"
+    },
+    {
+      "entity": "MAP data-acquisition system",
+      "label": "Instrument",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting summary - Software and code"
+    },
+    {
+      "entity": "Plexon Inc.",
+      "label": "Organization",
+      "sentence": "Data were recorded using the MAP data-acquisition system (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting summary - Software and code"
+    },
+    {
+      "entity": "Spike sorting",
+      "label": "Method",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting summary - Software and code"
+    },
+    {
+      "entity": "principal component analysis",
+      "label": "Method",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting summary - Software and code"
+    },
+    {
+      "entity": "Plexon offline sorter",
+      "label": "Software",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting summary - Software and code"
+    },
+    {
+      "entity": "Custom Python code",
+      "label": "Software",
+      "sentence": "Custom Python code was used for data analyses.",
+      "paper_location": "Reporting summary - Data analysis"
+    },
+    {
+      "entity": "Python",
+      "label": "Software",
+      "sentence": "Plots were generated with Python (version 3.8) and Seaborn (version 0.9.0).",
+      "paper_location": "Reporting summary - Data analysis"
+    },
+    {
+      "entity": "Seaborn",
+      "label": "Software",
+      "sentence": "Plots were generated with Python (version 3.8) and Seaborn (version 0.9.0).",
+      "paper_location": "Reporting summary - Data analysis"
+    },
+    {
+      "entity": "PyTorch",
+      "label": "Software",
+      "sentence": "Recurrent neural networks were trained using PyTorch (version 2.4).",
+      "paper_location": "Reporting summary - Data analysis"
+    },
+    {
+      "entity": "Recurrent neural networks",
+      "label": "ComputationalModel",
+      "sentence": "Recurrent neural networks were trained using PyTorch (version 2.4).",
+      "paper_location": "Reporting summary - Data analysis"
+    },
+    {
+      "entity": "GitHub",
+      "label": "SoftwareRepository",
+      "sentence": "The source code to reproduce results of this study is available on GitHub (https://github.com/engellab/latentcircuit) and Zenodo (https://zenodo.org/records/14020108).",
+      "paper_location": "Reporting summary - Data analysis"
+    },
+    {
+      "entity": "Zenodo",
+      "label": "SoftwareRepository",
+      "sentence": "The source code to reproduce results of this study is available on GitHub (https://github.com/engellab/latentcircuit) and Zenodo (https://zenodo.org/records/14020108).",
+      "paper_location": "Reporting summary - Data analysis"
+    }
+  ],
+  "key_terms": [
+    {
+      "term": "Latent circuit inference",
+      "sentence": "Latent circuit inference from heterogeneous neural responses during cognitive tasks",
+      "paper_location": "Title"
+    },
+    {
+      "term": "heterogeneous neural responses",
+      "sentence": "Latent circuit inference from heterogeneous neural responses during cognitive tasks",
+      "paper_location": "Title"
+    },
+    {
+      "term": "cognitive tasks",
+      "sentence": "Latent circuit inference from heterogeneous neural responses during cognitive tasks",
+      "paper_location": "Title"
+    },
+    {
+      "term": "low-dimensional recurrent connectivity",
+      "sentence": "We develop the latent circuit model, a dimensionality reduction approach in which task variables interact via low-dimensional recurrent connectivity to produce behavioral output.",
+      "paper_location": "Body"
+    },
+    {
+      "term": "context-dependent decision-making task",
+      "sentence": "We apply the latent circuit inference to recurrent neural networks trained to perform a context-dependent decision-making task and find a suppression mechanism in which contextual representations inhibit irrelevant sensory responses.",
+      "paper_location": "Body"
+    },
+    {
+      "term": "patterned connectivity perturbations",
+      "sentence": "We validate this mechanism by confirming the behavioral effects of patterned connectivity perturbations predicted by the latent circuit model.",
+      "paper_location": "Body"
+    },
+    {
+      "term": "causal interactions among task variables",
+      "sentence": "We show that incorporating causal interactions among task variables is critical for identifying behaviorally relevant computations from neural response data.",
+      "paper_location": "Body"
+    },
+    {
+      "term": "hand-crafted neural circuit models",
+      "sentence": "Traditionally, hand-crafted neural circuit models were used to pose specific mechanistic hypotheses about how excitation and inhibition between a few neural populations representing task variables control the flow of information from input to behavioral output [1][2][3][4][5][6][7][8][9][10][11] .",
+      "paper_location": "Body"
+    },
+    {
+      "term": "heterogeneous tuning",
+      "sentence": "Single neurons in areas such as the prefrontal cortex (PFC) show complex heterogeneous tuning to multiple task variables [16][17][18][19][20][21] , which presents a formidable challenge for identifying underlying circuit mechanisms.",
+      "paper_location": "Body"
+    },
+    {
+      "term": "correlation-based dimensionality reduction methods",
+      "sentence": "Using RNN perturbations, we show that dimensionality reduction methods that do not incorporate causal interactions among latent variables are biased toward uncovering behaviorally irrelevant representations.",
+      "paper_location": "Body"
+    },
+    {
+      "term": "low-dimensional circuit mechanism",
+      "sentence": "Our model fits neural responses with dynamics generated by a low-dimensional latent circuit, thereby directly testing the hypothesis that heterogeneous neural responses arise from a low-dimensional circuit mechanism.",
+      "paper_location": "Body"
+    },
+    {
+      "term": "inhibition of irrelevant sensory representations",
+      "sentence": "We applied latent circuit inference to RNNs optimized on a context-dependent decision-making task and found a circuit mechanism based on the inhibition of irrelevant sensory representations.",
+      "paper_location": "Body"
+    },
+    {
+      "term": "alternative stimulus-response mappings",
+      "sentence": "Context-dependent decision-making requires flexible trial-by-trial switching between alternative stimulus-response mappings.",
+      "paper_location": "Body"
+    },
+    {
+      "term": "inhibitory circuit mechanism",
+      "sentence": "By contrast, dimensionality reduction methods applied to PFC data or RNN activity show minimal suppression of irrelevant sensory responses 19,21 , seemingly invalidating the inhibitory circuit mechanism.",
+      "paper_location": "Body"
+    },
+    {
+      "term": "latent vector field",
+      "sentence": "Using the orthonormality of Q, we then derive the relationship: Q T V(Qx) = v(x), which asserts that the latent vector field v(x) describes dynamics of the highdimensional system in this invariant subspace.",
+      "paper_location": "Body"
+    },
+    {
+      "term": "rank-one connectivity perturbation",
+      "sentence": "Perturbing connection δ ij from node j to node i in the latent circuit maps onto rank-one connectivity perturbation Qδ ij Q T = q i q T j in the RNN.",
+      "paper_location": "Body"
+    },
+    {
+      "term": "embedding matrix Q",
+      "sentence": "where Q ∈ ℝ N×n is an orthonormal embedding matrix and n ≪ N.",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "term": "rectified linear (ReLU) activation function",
+      "sentence": "where f is a rectified linear (ReLU) activation function.",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "term": "orthonormal embedding matrix",
+      "sentence": "where Q ∈ ℝ N×n is an orthonormal embedding matrix and n ≪ N.",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "term": "low-dimensional subspace",
+      "sentence": "The latent circuit model captures task-related neural activity in the low-dimensional subspace spanned by the columns of Q, with dynamics within this subspace generated by the neural circuit Eq. ( 2).",
+      "paper_location": "Results - Latent circuit model"
+    },
+    {
+      "term": "latent connectivity",
+      "sentence": "The advantage of the mechanistic model for latent dynamics is that we can interpret the latent connectivity and relate it to the connectivity of the high-dimensional system.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "term": "piecewise-linear dynamical systems",
+      "sentence": "Using the fact that the vector fields of the RNN and latent circuit are piecewise-linear dynamical systems, we derive a relationship between their recurrent and input connectivity matrices (Methods):",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "term": "invariant subspace",
+      "sentence": "Rather, it is a weaker condition that the linear subspace defined by Q is an invariant subspace of the high-dimensional recurrent connectivity matrix.",
+      "paper_location": "Results - Interpreting latent connectivity"
+    },
+    {
+      "term": "latent connectivity structure",
+      "sentence": "This agreement confirms that the latent connectivity structure indeed exists in the RNN.",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "term": "suppression of irrelevant sensory representations",
+      "sentence": "The suppression of irrelevant sensory representations in RNN activity is consistent with the inhibitory mechanism revealed in the latent circuit connectivity w rec .",
+      "paper_location": "Results - Latent circuit for context-dependent decision-making"
+    },
+    {
+      "term": "space of latent circuit solutions",
+      "sentence": "To visualize the space of latent circuit solutions, we applied a principal component analysis to the flattened connectivity matrices w rec and projected the data onto the first two principal components, which accounted for 42% of total variance (Fig. 5a).",
+      "paper_location": "Results - Space of latent circuit mechanisms"
+    },
+    {
+      "term": "dynamical selection vector mechanism",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses 19 .",
+      "paper_location": "Results - Space of latent circuit mechanisms"
+    },
+    {
+      "term": "inhibitory suppression mechanism",
+      "sentence": "The inhibitory suppression mechanism that we consistently found across RNNs may seem distinct from a dynamical selection vector mechanism previously identified in RNNs trained on a similar task, which apparently does not require suppression of irrelevant sensory responses 19 .",
+      "paper_location": "Results - Space of latent circuit mechanisms"
+    },
+    {
+      "term": "error trials",
+      "sentence": "Although we cannot directly assess the behavioral relevance of different representations in the PFC via perturbations, we can test the inhibitory suppression mechanism using neural activity on error trials.",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "term": "incongruent trials",
+      "sentence": "We tested this prediction on incongruent trials on which the relevant and irrelevant stimuli point to opposite choices (Methods) because the errors on incongruent trials are more likely to result from a failure of the contextual mechanism than other factors (for example, attention lapse).",
+      "paper_location": "Results - Representations of irrelevant stimuli in the PFC"
+    },
+    {
+      "term": "behavioral relevance of representations",
+      "sentence": "To further test for behavioral relevance of representations identified by different dimensionality reduction methods, we again turned to RNNs in which we can measure behavioral effects of arbitrary activity perturbations.",
+      "paper_location": "Results - Behavioral relevance of low-dimensional representations"
+    },
+    {
+      "term": "decoder compromises behavioral relevance for decoding accuracy",
+      "sentence": "These perturbations reveal that the decoder compromises behavioral relevance for decoding accuracy.",
+      "paper_location": "Figure 7"
+    },
+    {
+      "term": "mean squared error loss function",
+      "sentence": "We fit the latent circuit model Eqs. (1)-(3) to neural response data y by minimizing the mean squared error loss function,",
+      "paper_location": "Methods - Fitting a latent circuit model"
+    },
+    {
+      "term": "nonconvex optimization",
+      "sentence": "This ensemble of latent circuit models for a single RNN has variable fit quality because many optimization runs do not converge to the optimal solution (which is typical for nonconvex optimization).",
+      "paper_location": "Methods - Fitting a latent circuit model"
+    },
+    {
+      "term": "permutation test",
+      "sentence": "To determine whether the inferred latent circuit connectivity significantly depends on neural response data beyond the constraints imposed by the task alone, we performed a permutation test (Extended Data Figs. 1, 7c and 8c), which proceeds in three steps.",
+      "paper_location": "Methods - Testing the dependence of latent connectivity on neural responses"
+    },
+    {
+      "term": "time-discretized RNNs",
+      "sentence": "We simulate dynamics of time-discretized RNNs using the general framework for modeling cognitive tasks 22 .",
+      "paper_location": "Methods - RNN simulations"
+    },
+    {
+      "term": "Dale's law",
+      "sentence": "The recurrent matrix is constrained to satisfy Dale's law with 80% excitatory units and 20% inhibitory units.",
+      "paper_location": "Methods - RNN simulations"
+    },
+    {
+      "term": "motion and color coherence",
+      "sentence": "The strength of motion and color stimuli varies from trial to trial as quantified by the motion and color coherence.",
+      "paper_location": "Methods - Context-dependent decision-making task"
+    },
+    {
+      "term": "spike sorting",
+      "sentence": "Spike sorting was conducted by clustering based on principal component analysis using the Plexon offline sorter (Plexon Inc., Dallas, TX).",
+      "paper_location": "Reporting summary - Software and code"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/evaluation/ner/direct_llm_call/output/multiscale_spatial_transcriptomic/neuroscience_ner_openai-gpt-5.5_20260702_192655.json b/evaluation/ner/direct_llm_call/output/multiscale_spatial_transcriptomic/neuroscience_ner_openai-gpt-5.5_20260702_192655.json
new file mode 100644
index 0000000..6381522
--- /dev/null
+++ b/evaluation/ner/direct_llm_call/output/multiscale_spatial_transcriptomic/neuroscience_ner_openai-gpt-5.5_20260702_192655.json
@@ -0,0 +1,19 @@
+{
+  "raw_output": "{\n  \"source_metadata\": {\n    \"paper_title\": \"Multiscale Spatial Transcriptomic Atlas of Human Basal Ganglia Cell-Type and Cellular Community Organization\",\n    \"doi\": \"10.64898/2025.12.02.691876\",\n    \"source_path\": \"evaluation/ner/multiscale_spatial_transcriptomic/2025.12.02.691876v1.full.pdf\"\n  },\n  \"entities\": [\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"<term>MERFISH+</term>\",\n      \"paper_location\": \"Keywords\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"<term>Stereo-seq</term>\",\n      \"paper_location\": \"Keywords\"\n    },\n    {\n      \"entity\": \"human basal ganglia\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"• Our multi-centimeter scale spatial taxonomy identifies the precise locations of 60 neuronal and glial cell types of human basal ganglia.\",\n      \"paper_location\": \"Highlights\"\n    },\n    {\n      \"entity\": \"neuronal and glial cell types\",\n      \"label\": \"CellType\",\n      \"sentence\": \"• Our multi-centimeter scale spatial taxonomy identifies the precise locations of 60 neuronal and glial cell types of human basal ganglia.\",\n      \"paper_location\": \"Highlights\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"• MERFISH+ and Stereo-seq platforms map consistent spatial modules that align with classical neuroanatomical nuclei.\",\n      \"paper_location\": \"Highlights\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"• MERFISH+ and Stereo-seq platforms map consistent spatial modules that align with classical neuroanatomical nuclei.\",\n      \"paper_location\": \"Highlights\"\n    },\n    {\n      \"entity\": \"D1D2 hybrid MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"• D1D2 hybrid MSNs and primate-expanded interneurons show regional and domain specific organization\",\n      \"paper_location\": \"Highlights\"\n    },\n    {\n      \"entity\": \"primate-expanded interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"• D1D2 hybrid MSNs and primate-expanded interneurons show regional and domain specific organization\",\n      \"paper_location\": \"Highlights\"\n    },\n    {\n      \"entity\": \"Subcellular RNA localization\",\n      \"label\": \"Method\",\n      \"sentence\": \"• Subcellular RNA localization reports soma morphology and projection-inferred signatures.\",\n      \"paper_location\": \"Highlights\"\n    },\n    {\n      \"entity\": \"soma morphology\",\n      \"label\": \"Measurement\",\n      \"sentence\": \"• Subcellular RNA localization reports soma morphology and projection-inferred signatures.\",\n      \"paper_location\": \"Highlights\"\n    },\n    {\n      \"entity\": \"projection-inferred signatures\",\n      \"label\": \"Measurement\",\n      \"sentence\": \"• Subcellular RNA localization reports soma morphology and projection-inferred signatures.\",\n      \"paper_location\": \"Highlights\"\n    },\n    {\n      \"entity\": \"subcellular-resolution spatial transcriptomic atlas\",\n      \"label\": \"Method\",\n      \"sentence\": \"We generated a multi-region, subcellular-resolution spatial transcriptomic atlas of the human basal ganglia by integrating MERFISH+ and Stereo-seq across four neurotypical donors.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"human basal ganglia\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"We generated a multi-region, subcellular-resolution spatial transcriptomic atlas of the human basal ganglia by integrating MERFISH+ and Stereo-seq across four neurotypical donors.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"We generated a multi-region, subcellular-resolution spatial transcriptomic atlas of the human basal ganglia by integrating MERFISH+ and Stereo-seq across four neurotypical donors.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"We generated a multi-region, subcellular-resolution spatial transcriptomic atlas of the human basal ganglia by integrating MERFISH+ and Stereo-seq across four neurotypical donors.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"caudate\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"These datasets profiled ~7 million cells spanning the caudate, putamen, nucleus accumbens, and globus pallidus, resolving 60 transcriptionally distinct cell types.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"putamen\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"These datasets profiled ~7 million cells spanning the caudate, putamen, nucleus accumbens, and globus pallidus, resolving 60 transcriptionally distinct cell types.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"nucleus accumbens\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"These datasets profiled ~7 million cells spanning the caudate, putamen, nucleus accumbens, and globus pallidus, resolving 60 transcriptionally distinct cell types.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"globus pallidus\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"These datasets profiled ~7 million cells spanning the caudate, putamen, nucleus accumbens, and globus pallidus, resolving 60 transcriptionally distinct cell types.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"medium-spiny-neuron cell types\",\n      \"label\": \"CellType\",\n      \"sentence\": \"We show region-selective, molecular and spatial diversification of medium-spiny-neuron cell types and multiple non-neuronal populations with distinct molecular identities and spatial localizations.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"non-neuronal populations\",\n      \"label\": \"CellType\",\n      \"sentence\": \"We show region-selective, molecular and spatial diversification of medium-spiny-neuron cell types and multiple non-neuronal populations with distinct molecular identities and spatial localizations.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"Subcellular RNA localization\",\n      \"label\": \"Method\",\n      \"sentence\": \"Subcellular RNA localization captures somatic size and projectioninferred signatures that reflect direct and indirect pathway topology.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"direct and indirect pathway topology\",\n      \"label\": \"NeuralCircuit\",\n      \"sentence\": \"Subcellular RNA localization captures somatic size and projectioninferred signatures that reflect direct and indirect pathway topology.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"astrocytes\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Cellular community analyses reveal the enrichment of sub-clusters of astrocytes and oligodendrocytes at striosome-matrix borders, while primate-expanded interneurons are confined to matrix territories.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"oligodendrocytes\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Cellular community analyses reveal the enrichment of sub-clusters of astrocytes and oligodendrocytes at striosome-matrix borders, while primate-expanded interneurons are confined to matrix territories.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"striosome-matrix borders\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"Cellular community analyses reveal the enrichment of sub-clusters of astrocytes and oligodendrocytes at striosome-matrix borders, while primate-expanded interneurons are confined to matrix territories.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"primate-expanded interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Cellular community analyses reveal the enrichment of sub-clusters of astrocytes and oligodendrocytes at striosome-matrix borders, while primate-expanded interneurons are confined to matrix territories.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"striosome-matrix organization\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"Cross-species mapping uncovers orthologous striosome-matrix organization and conserved dorsolateral-ventromedial gene expression gradients.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"dorsolateral-ventromedial gene expression gradients\",\n      \"label\": \"Measurement\",\n      \"sentence\": \"Cross-species mapping uncovers orthologous striosome-matrix organization and conserved dorsolateral-ventromedial gene expression gradients.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"human basal ganglia\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"This atlas provides a foundational molecular and spatial framework for studying human basal ganglia architecture, offering a multi-centimeter scale resource that links cell types, spatial architecture, and subcellular transcript topography across multiple nuclei.\",\n      \"paper_location\": \"SUMMARY\"\n    },\n    {\n      \"entity\": \"human basal ganglia\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"The human basal ganglia is a critical brain structure involved in motor control, learning and emotion.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"basal ganglia\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"The human basal ganglia is a critical brain structure involved in motor control, learning and emotion.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"striatum\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"It is comprised of distinct anatomical nuclei / subdivisions including the striatum (caudate nucleus and putamen), the globus pallidus (external and internal segments, GPe and GPi), and the nucleus accumbens (NAc) <ref type=\\\"bibr\\\" target=\\\"#b0\\\">1</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"caudate nucleus\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"It is comprised of distinct anatomical nuclei / subdivisions including the striatum (caudate nucleus and putamen), the globus pallidus (external and internal segments, GPe and GPi), and the nucleus accumbens (NAc) <ref type=\\\"bibr\\\" target=\\\"#b0\\\">1</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"putamen\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"It is comprised of distinct anatomical nuclei / subdivisions including the striatum (caudate nucleus and putamen), the globus pallidus (external and internal segments, GPe and GPi), and the nucleus accumbens (NAc) <ref type=\\\"bibr\\\" target=\\\"#b0\\\">1</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"globus pallidus\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"It is comprised of distinct anatomical nuclei / subdivisions including the striatum (caudate nucleus and putamen), the globus pallidus (external and internal segments, GPe and GPi), and the nucleus accumbens (NAc) <ref type=\\\"bibr\\\" target=\\\"#b0\\\">1</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"GPe\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"It is comprised of distinct anatomical nuclei / subdivisions including the striatum (caudate nucleus and putamen), the globus pallidus (external and internal segments, GPe and GPi), and the nucleus accumbens (NAc) <ref type=\\\"bibr\\\" target=\\\"#b0\\\">1</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"GPi\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"It is comprised of distinct anatomical nuclei / subdivisions including the striatum (caudate nucleus and putamen), the globus pallidus (external and internal segments, GPe and GPi), and the nucleus accumbens (NAc) <ref type=\\\"bibr\\\" target=\\\"#b0\\\">1</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"nucleus accumbens\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"It is comprised of distinct anatomical nuclei / subdivisions including the striatum (caudate nucleus and putamen), the globus pallidus (external and internal segments, GPe and GPi), and the nucleus accumbens (NAc) <ref type=\\\"bibr\\\" target=\\\"#b0\\\">1</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"NAc\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"It is comprised of distinct anatomical nuclei / subdivisions including the striatum (caudate nucleus and putamen), the globus pallidus (external and internal segments, GPe and GPi), and the nucleus accumbens (NAc) <ref type=\\\"bibr\\\" target=\\\"#b0\\\">1</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"GABAergic medium spiny neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"striatum\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"D1-type MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"dopamine receptor D1\",\n      \"label\": \"Protein\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"DRD1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"tachykinin precursor 1\",\n      \"label\": \"Protein\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"TAC1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"dynorphin\",\n      \"label\": \"Neuropeptide\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"D2-type MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"dopamine receptor D2\",\n      \"label\": \"Protein\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"DRD2\",\n      \"label\": \"Gene\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"proenkephalin\",\n      \"label\": \"Neuropeptide\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"PENK\",\n      \"label\": \"Gene\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"enkephalin\",\n      \"label\": \"Neuropeptide\",\n      \"sentence\": \"The literature from both animal models and humans shows that GABAergic medium spiny neurons (MSNs) are the dominant neuronal cell type in the striatum (85-95%) <ref type=\\\"bibr\\\" target=\\\"#b1\\\">[2]</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">[3]</ref><ref type=\\\"bibr\\\" target=\\\"#b3\\\">[4]</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">[5]</ref> ; MSNs are divided into two major populations: D1-type MSNs, which express the dopamine receptor D1 (DRD1), tachykinin precursor 1 (TAC1), and dynorphin; and D2-type MSNs, which express the dopamine receptor D2 (DRD2), proenkephalin (PENK), and enkephalin <ref type=\\\"bibr\\\" target=\\\"#b1\\\">2,</ref><ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"D1\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The D1 and D2 populations are spatially organized into interdigitated striosome and matrix compartments that form the structural foundation of basal ganglia function <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b6\\\">7,</ref><ref type=\\\"bibr\\\" target=\\\"#b7\\\">8</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"D2\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The D1 and D2 populations are spatially organized into interdigitated striosome and matrix compartments that form the structural foundation of basal ganglia function <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b6\\\">7,</ref><ref type=\\\"bibr\\\" target=\\\"#b7\\\">8</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"striosome\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"The D1 and D2 populations are spatially organized into interdigitated striosome and matrix compartments that form the structural foundation of basal ganglia function <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b6\\\">7,</ref><ref type=\\\"bibr\\\" target=\\\"#b7\\\">8</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"matrix compartments\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"The D1 and D2 populations are spatially organized into interdigitated striosome and matrix compartments that form the structural foundation of basal ganglia function <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b6\\\">7,</ref><ref type=\\\"bibr\\\" target=\\\"#b7\\\">8</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"single-nucleus RNA-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"Further studies, including those using singlenucleus RNA-seq in animal and human brains have revealed additional D1/D2 neuronal diversity <ref type=\\\"bibr\\\" target=\\\"#b8\\\">9,</ref><ref type=\\\"bibr\\\" target=\\\"#b9\\\">10</ref> , and identified multiple classes of interneurons, including cholinergic (choline acetyltransferase-positive; ChAT+), parvalbumin-positive (PV+), somatostatinpositive (SST+), and neuropeptide Y/tyrosine hydroxylase-positive (NPY+/TH+) subtypes <ref type=\\\"bibr\\\" target=\\\"#b9\\\">[10]</ref><ref type=\\\"bibr\\\" target=\\\"#b10\\\">[11]</ref><ref type=\\\"bibr\\\" target=\\\"#b11\\\">[12]</ref><ref type=\\\"bibr\\\" target=\\\"#b12\\\">[13]</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">[14]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Further studies, including those using singlenucleus RNA-seq in animal and human brains have revealed additional D1/D2 neuronal diversity <ref type=\\\"bibr\\\" target=\\\"#b8\\\">9,</ref><ref type=\\\"bibr\\\" target=\\\"#b9\\\">10</ref> , and identified multiple classes of interneurons, including cholinergic (choline acetyltransferase-positive; ChAT+), parvalbumin-positive (PV+), somatostatinpositive (SST+), and neuropeptide Y/tyrosine hydroxylase-positive (NPY+/TH+) subtypes <ref type=\\\"bibr\\\" target=\\\"#b9\\\">[10]</ref><ref type=\\\"bibr\\\" target=\\\"#b10\\\">[11]</ref><ref type=\\\"bibr\\\" target=\\\"#b11\\\">[12]</ref><ref type=\\\"bibr\\\" target=\\\"#b12\\\">[13]</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">[14]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"cholinergic\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Further studies, including those using singlenucleus RNA-seq in animal and human brains have revealed additional D1/D2 neuronal diversity <ref type=\\\"bibr\\\" target=\\\"#b8\\\">9,</ref><ref type=\\\"bibr\\\" target=\\\"#b9\\\">10</ref> , and identified multiple classes of interneurons, including cholinergic (choline acetyltransferase-positive; ChAT+), parvalbumin-positive (PV+), somatostatinpositive (SST+), and neuropeptide Y/tyrosine hydroxylase-positive (NPY+/TH+) subtypes <ref type=\\\"bibr\\\" target=\\\"#b9\\\">[10]</ref><ref type=\\\"bibr\\\" target=\\\"#b10\\\">[11]</ref><ref type=\\\"bibr\\\" target=\\\"#b11\\\">[12]</ref><ref type=\\\"bibr\\\" target=\\\"#b12\\\">[13]</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">[14]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"choline acetyltransferase-positive\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Further studies, including those using singlenucleus RNA-seq in animal and human brains have revealed additional D1/D2 neuronal diversity <ref type=\\\"bibr\\\" target=\\\"#b8\\\">9,</ref><ref type=\\\"bibr\\\" target=\\\"#b9\\\">10</ref> , and identified multiple classes of interneurons, including cholinergic (choline acetyltransferase-positive; ChAT+), parvalbumin-positive (PV+), somatostatinpositive (SST+), and neuropeptide Y/tyrosine hydroxylase-positive (NPY+/TH+) subtypes <ref type=\\\"bibr\\\" target=\\\"#b9\\\">[10]</ref><ref type=\\\"bibr\\\" target=\\\"#b10\\\">[11]</ref><ref type=\\\"bibr\\\" target=\\\"#b11\\\">[12]</ref><ref type=\\\"bibr\\\" target=\\\"#b12\\\">[13]</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">[14]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"ChAT+\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Further studies, including those using singlenucleus RNA-seq in animal and human brains have revealed additional D1/D2 neuronal diversity <ref type=\\\"bibr\\\" target=\\\"#b8\\\">9,</ref><ref type=\\\"bibr\\\" target=\\\"#b9\\\">10</ref> , and identified multiple classes of interneurons, including cholinergic (choline acetyltransferase-positive; ChAT+), parvalbumin-positive (PV+), somatostatinpositive (SST+), and neuropeptide Y/tyrosine hydroxylase-positive (NPY+/TH+) subtypes <ref type=\\\"bibr\\\" target=\\\"#b9\\\">[10]</ref><ref type=\\\"bibr\\\" target=\\\"#b10\\\">[11]</ref><ref type=\\\"bibr\\\" target=\\\"#b11\\\">[12]</ref><ref type=\\\"bibr\\\" target=\\\"#b12\\\">[13]</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">[14]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"parvalbumin-positive\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Further studies, including those using singlenucleus RNA-seq in animal and human brains have revealed additional D1/D2 neuronal diversity <ref type=\\\"bibr\\\" target=\\\"#b8\\\">9,</ref><ref type=\\\"bibr\\\" target=\\\"#b9\\\">10</ref> , and identified multiple classes of interneurons, including cholinergic (choline acetyltransferase-positive; ChAT+), parvalbumin-positive (PV+), somatostatinpositive (SST+), and neuropeptide Y/tyrosine hydroxylase-positive (NPY+/TH+) subtypes <ref type=\\\"bibr\\\" target=\\\"#b9\\\">[10]</ref><ref type=\\\"bibr\\\" target=\\\"#b10\\\">[11]</ref><ref type=\\\"bibr\\\" target=\\\"#b11\\\">[12]</ref><ref type=\\\"bibr\\\" target=\\\"#b12\\\">[13]</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">[14]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"PV+\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Further studies, including those using singlenucleus RNA-seq in animal and human brains have revealed additional D1/D2 neuronal diversity <ref type=\\\"bibr\\\" target=\\\"#b8\\\">9,</ref><ref type=\\\"bibr\\\" target=\\\"#b9\\\">10</ref> , and identified multiple classes of interneurons, including cholinergic (choline acetyltransferase-positive; ChAT+), parvalbumin-positive (PV+), somatostatinpositive (SST+), and neuropeptide Y/tyrosine hydroxylase-positive (NPY+/TH+) subtypes <ref type=\\\"bibr\\\" target=\\\"#b9\\\">[10]</ref><ref type=\\\"bibr\\\" target=\\\"#b10\\\">[11]</ref><ref type=\\\"bibr\\\" target=\\\"#b11\\\">[12]</ref><ref type=\\\"bibr\\\" target=\\\"#b12\\\">[13]</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">[14]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"somatostatinpositive\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Further studies, including those using singlenucleus RNA-seq in animal and human brains have revealed additional D1/D2 neuronal diversity <ref type=\\\"bibr\\\" target=\\\"#b8\\\">9,</ref><ref type=\\\"bibr\\\" target=\\\"#b9\\\">10</ref> , and identified multiple classes of interneurons, including cholinergic (choline acetyltransferase-positive; ChAT+), parvalbumin-positive (PV+), somatostatinpositive (SST+), and neuropeptide Y/tyrosine hydroxylase-positive (NPY+/TH+) subtypes <ref type=\\\"bibr\\\" target=\\\"#b9\\\">[10]</ref><ref type=\\\"bibr\\\" target=\\\"#b10\\\">[11]</ref><ref type=\\\"bibr\\\" target=\\\"#b11\\\">[12]</ref><ref type=\\\"bibr\\\" target=\\\"#b12\\\">[13]</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">[14]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"SST+\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Further studies, including those using singlenucleus RNA-seq in animal and human brains have revealed additional D1/D2 neuronal diversity <ref type=\\\"bibr\\\" target=\\\"#b8\\\">9,</ref><ref type=\\\"bibr\\\" target=\\\"#b9\\\">10</ref> , and identified multiple classes of interneurons, including cholinergic (choline acetyltransferase-positive; ChAT+), parvalbumin-positive (PV+), somatostatinpositive (SST+), and neuropeptide Y/tyrosine hydroxylase-positive (NPY+/TH+) subtypes <ref type=\\\"bibr\\\" target=\\\"#b9\\\">[10]</ref><ref type=\\\"bibr\\\" target=\\\"#b10\\\">[11]</ref><ref type=\\\"bibr\\\" target=\\\"#b11\\\">[12]</ref><ref type=\\\"bibr\\\" target=\\\"#b12\\\">[13]</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">[14]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"neuropeptide Y/tyrosine hydroxylase-positive\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Further studies, including those using singlenucleus RNA-seq in animal and human brains have revealed additional D1/D2 neuronal diversity <ref type=\\\"bibr\\\" target=\\\"#b8\\\">9,</ref><ref type=\\\"bibr\\\" target=\\\"#b9\\\">10</ref> , and identified multiple classes of interneurons, including cholinergic (choline acetyltransferase-positive; ChAT+), parvalbumin-positive (PV+), somatostatinpositive (SST+), and neuropeptide Y/tyrosine hydroxylase-positive (NPY+/TH+) subtypes <ref type=\\\"bibr\\\" target=\\\"#b9\\\">[10]</ref><ref type=\\\"bibr\\\" target=\\\"#b10\\\">[11]</ref><ref type=\\\"bibr\\\" target=\\\"#b11\\\">[12]</ref><ref type=\\\"bibr\\\" target=\\\"#b12\\\">[13]</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">[14]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"NPY+/TH+\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Further studies, including those using singlenucleus RNA-seq in animal and human brains have revealed additional D1/D2 neuronal diversity <ref type=\\\"bibr\\\" target=\\\"#b8\\\">9,</ref><ref type=\\\"bibr\\\" target=\\\"#b9\\\">10</ref> , and identified multiple classes of interneurons, including cholinergic (choline acetyltransferase-positive; ChAT+), parvalbumin-positive (PV+), somatostatinpositive (SST+), and neuropeptide Y/tyrosine hydroxylase-positive (NPY+/TH+) subtypes <ref type=\\\"bibr\\\" target=\\\"#b9\\\">[10]</ref><ref type=\\\"bibr\\\" target=\\\"#b10\\\">[11]</ref><ref type=\\\"bibr\\\" target=\\\"#b11\\\">[12]</ref><ref type=\\\"bibr\\\" target=\\\"#b12\\\">[13]</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">[14]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"nucleus accumbens\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"The nucleus accumbens, located at the ventral extension of the striatum, contains both D1 and D2 MSNs and is further divided into core and shell subregions.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"striatum\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"The nucleus accumbens, located at the ventral extension of the striatum, contains both D1 and D2 MSNs and is further divided into core and shell subregions.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"D1\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The nucleus accumbens, located at the ventral extension of the striatum, contains both D1 and D2 MSNs and is further divided into core and shell subregions.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"D2 MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The nucleus accumbens, located at the ventral extension of the striatum, contains both D1 and D2 MSNs and is further divided into core and shell subregions.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"dorsal striatum\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"These subregions mirror the cellular composition of the dorsal striatum, including diverse interneuron classes <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b14\\\">15</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"interneuron classes\",\n      \"label\": \"CellType\",\n      \"sentence\": \"These subregions mirror the cellular composition of the dorsal striatum, including diverse interneuron classes <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b14\\\">15</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"globus pallidus\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"The globus pallidus consists of two principal divisions: the external segment (GPe) and the internal segment (GPi) <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b15\\\">16</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"external segment\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"The globus pallidus consists of two principal divisions: the external segment (GPe) and the internal segment (GPi) <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b15\\\">16</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"GPe\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"The globus pallidus consists of two principal divisions: the external segment (GPe) and the internal segment (GPi) <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b15\\\">16</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"internal segment\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"The globus pallidus consists of two principal divisions: the external segment (GPe) and the internal segment (GPi) <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b15\\\">16</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"GPi\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"The globus pallidus consists of two principal divisions: the external segment (GPe) and the internal segment (GPi) <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b15\\\">16</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"PV+ GABAergic projection neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Both regions are dominated by PV+ GABAergic projection neurons <ref type=\\\"bibr\\\" target=\\\"#b16\\\">17</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"GPe\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Within GPe, prototypic neurons express PV, LIM homeobox 6 (LHX6), and NK2 homeobox 1 (NKX2-1) <ref type=\\\"bibr\\\" target=\\\"#b17\\\">18,</ref><ref type=\\\"bibr\\\" target=\\\"#b18\\\">19</ref> ; arkypallidal neurons express neuronal PAS domain protein 1 (NPAS1) and forkhead box protein P2 (FOXP2) <ref type=\\\"bibr\\\" target=\\\"#b19\\\">20,</ref><ref type=\\\"bibr\\\" target=\\\"#b20\\\">21</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"prototypic neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Within GPe, prototypic neurons express PV, LIM homeobox 6 (LHX6), and NK2 homeobox 1 (NKX2-1) <ref type=\\\"bibr\\\" target=\\\"#b17\\\">18,</ref><ref type=\\\"bibr\\\" target=\\\"#b18\\\">19</ref> ; arkypallidal neurons express neuronal PAS domain protein 1 (NPAS1) and forkhead box protein P2 (FOXP2) <ref type=\\\"bibr\\\" target=\\\"#b19\\\">20,</ref><ref type=\\\"bibr\\\" target=\\\"#b20\\\">21</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"PV\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Within GPe, prototypic neurons express PV, LIM homeobox 6 (LHX6), and NK2 homeobox 1 (NKX2-1) <ref type=\\\"bibr\\\" target=\\\"#b17\\\">18,</ref><ref type=\\\"bibr\\\" target=\\\"#b18\\\">19</ref> ; arkypallidal neurons express neuronal PAS domain protein 1 (NPAS1) and forkhead box protein P2 (FOXP2) <ref type=\\\"bibr\\\" target=\\\"#b19\\\">20,</ref><ref type=\\\"bibr\\\" target=\\\"#b20\\\">21</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"LIM homeobox 6\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Within GPe, prototypic neurons express PV, LIM homeobox 6 (LHX6), and NK2 homeobox 1 (NKX2-1) <ref type=\\\"bibr\\\" target=\\\"#b17\\\">18,</ref><ref type=\\\"bibr\\\" target=\\\"#b18\\\">19</ref> ; arkypallidal neurons express neuronal PAS domain protein 1 (NPAS1) and forkhead box protein P2 (FOXP2) <ref type=\\\"bibr\\\" target=\\\"#b19\\\">20,</ref><ref type=\\\"bibr\\\" target=\\\"#b20\\\">21</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"LHX6\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Within GPe, prototypic neurons express PV, LIM homeobox 6 (LHX6), and NK2 homeobox 1 (NKX2-1) <ref type=\\\"bibr\\\" target=\\\"#b17\\\">18,</ref><ref type=\\\"bibr\\\" target=\\\"#b18\\\">19</ref> ; arkypallidal neurons express neuronal PAS domain protein 1 (NPAS1) and forkhead box protein P2 (FOXP2) <ref type=\\\"bibr\\\" target=\\\"#b19\\\">20,</ref><ref type=\\\"bibr\\\" target=\\\"#b20\\\">21</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"NK2 homeobox 1\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Within GPe, prototypic neurons express PV, LIM homeobox 6 (LHX6), and NK2 homeobox 1 (NKX2-1) <ref type=\\\"bibr\\\" target=\\\"#b17\\\">18,</ref><ref type=\\\"bibr\\\" target=\\\"#b18\\\">19</ref> ; arkypallidal neurons express neuronal PAS domain protein 1 (NPAS1) and forkhead box protein P2 (FOXP2) <ref type=\\\"bibr\\\" target=\\\"#b19\\\">20,</ref><ref type=\\\"bibr\\\" target=\\\"#b20\\\">21</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"NKX2-1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Within GPe, prototypic neurons express PV, LIM homeobox 6 (LHX6), and NK2 homeobox 1 (NKX2-1) <ref type=\\\"bibr\\\" target=\\\"#b17\\\">18,</ref><ref type=\\\"bibr\\\" target=\\\"#b18\\\">19</ref> ; arkypallidal neurons express neuronal PAS domain protein 1 (NPAS1) and forkhead box protein P2 (FOXP2) <ref type=\\\"bibr\\\" target=\\\"#b19\\\">20,</ref><ref type=\\\"bibr\\\" target=\\\"#b20\\\">21</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"arkypallidal neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Within GPe, prototypic neurons express PV, LIM homeobox 6 (LHX6), and NK2 homeobox 1 (NKX2-1) <ref type=\\\"bibr\\\" target=\\\"#b17\\\">18,</ref><ref type=\\\"bibr\\\" target=\\\"#b18\\\">19</ref> ; arkypallidal neurons express neuronal PAS domain protein 1 (NPAS1) and forkhead box protein P2 (FOXP2) <ref type=\\\"bibr\\\" target=\\\"#b19\\\">20,</ref><ref type=\\\"bibr\\\" target=\\\"#b20\\\">21</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"neuronal PAS domain protein 1\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Within GPe, prototypic neurons express PV, LIM homeobox 6 (LHX6), and NK2 homeobox 1 (NKX2-1) <ref type=\\\"bibr\\\" target=\\\"#b17\\\">18,</ref><ref type=\\\"bibr\\\" target=\\\"#b18\\\">19</ref> ; arkypallidal neurons express neuronal PAS domain protein 1 (NPAS1) and forkhead box protein P2 (FOXP2) <ref type=\\\"bibr\\\" target=\\\"#b19\\\">20,</ref><ref type=\\\"bibr\\\" target=\\\"#b20\\\">21</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"NPAS1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Within GPe, prototypic neurons express PV, LIM homeobox 6 (LHX6), and NK2 homeobox 1 (NKX2-1) <ref type=\\\"bibr\\\" target=\\\"#b17\\\">18,</ref><ref type=\\\"bibr\\\" target=\\\"#b18\\\">19</ref> ; arkypallidal neurons express neuronal PAS domain protein 1 (NPAS1) and forkhead box protein P2 (FOXP2) <ref type=\\\"bibr\\\" target=\\\"#b19\\\">20,</ref><ref type=\\\"bibr\\\" target=\\\"#b20\\\">21</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"forkhead box protein P2\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Within GPe, prototypic neurons express PV, LIM homeobox 6 (LHX6), and NK2 homeobox 1 (NKX2-1) <ref type=\\\"bibr\\\" target=\\\"#b17\\\">18,</ref><ref type=\\\"bibr\\\" target=\\\"#b18\\\">19</ref> ; arkypallidal neurons express neuronal PAS domain protein 1 (NPAS1) and forkhead box protein P2 (FOXP2) <ref type=\\\"bibr\\\" target=\\\"#b19\\\">20,</ref><ref type=\\\"bibr\\\" target=\\\"#b20\\\">21</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"FOXP2\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Within GPe, prototypic neurons express PV, LIM homeobox 6 (LHX6), and NK2 homeobox 1 (NKX2-1) <ref type=\\\"bibr\\\" target=\\\"#b17\\\">18,</ref><ref type=\\\"bibr\\\" target=\\\"#b18\\\">19</ref> ; arkypallidal neurons express neuronal PAS domain protein 1 (NPAS1) and forkhead box protein P2 (FOXP2) <ref type=\\\"bibr\\\" target=\\\"#b19\\\">20,</ref><ref type=\\\"bibr\\\" target=\\\"#b20\\\">21</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"GPi\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"GPi contains PV+/LHX6+ projection neurons that form the principal output pathway of the basal ganglia <ref type=\\\"bibr\\\" target=\\\"#b15\\\">16,</ref><ref type=\\\"bibr\\\" target=\\\"#b16\\\">17</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"PV+/LHX6+ projection neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"GPi contains PV+/LHX6+ projection neurons that form the principal output pathway of the basal ganglia <ref type=\\\"bibr\\\" target=\\\"#b15\\\">16,</ref><ref type=\\\"bibr\\\" target=\\\"#b16\\\">17</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"principal output pathway\",\n      \"label\": \"NeuralCircuit\",\n      \"sentence\": \"GPi contains PV+/LHX6+ projection neurons that form the principal output pathway of the basal ganglia <ref type=\\\"bibr\\\" target=\\\"#b15\\\">16,</ref><ref type=\\\"bibr\\\" target=\\\"#b16\\\">17</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"single-cell RNA sequencing\",\n      \"label\": \"Method\",\n      \"sentence\": \"Traditional single-cell RNA sequencing provides molecular resolution but lacks spatial context, limiting its ability to capture circuit-level organization.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"spatial transcriptomic technologies\",\n      \"label\": \"Method\",\n      \"sentence\": \"Emerging spatial transcriptomic technologies map gene expression directly within intact tissue sections that retain the spatial relationships of cell types and structures <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">14,</ref><ref type=\\\"bibr\\\" target=\\\"#b21\\\">[22]</ref><ref type=\\\"bibr\\\" target=\\\"#b22\\\">[23]</ref><ref type=\\\"bibr\\\" target=\\\"#b23\\\">[24]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"gene expression\",\n      \"label\": \"Measurement\",\n      \"sentence\": \"Emerging spatial transcriptomic technologies map gene expression directly within intact tissue sections that retain the spatial relationships of cell types and structures <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">14,</ref><ref type=\\\"bibr\\\" target=\\\"#b21\\\">[22]</ref><ref type=\\\"bibr\\\" target=\\\"#b22\\\">[23]</ref><ref type=\\\"bibr\\\" target=\\\"#b23\\\">[24]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"multiplexed fluorescent in situ hybridization\",\n      \"label\": \"Method\",\n      \"sentence\": \"These methods include probe-based multiplexed fluorescent in situ hybridization (FISH) and next-generation sequencing (NGS)-based spatially resolved transcriptomics, each with complementary strengths in resolution, transcriptome coverage, and throughput <ref type=\\\"bibr\\\" target=\\\"#b22\\\">23,</ref><ref type=\\\"bibr\\\" target=\\\"#b24\\\">25,</ref><ref type=\\\"bibr\\\" target=\\\"#b25\\\">26</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"FISH\",\n      \"label\": \"Method\",\n      \"sentence\": \"These methods include probe-based multiplexed fluorescent in situ hybridization (FISH) and next-generation sequencing (NGS)-based spatially resolved transcriptomics, each with complementary strengths in resolution, transcriptome coverage, and throughput <ref type=\\\"bibr\\\" target=\\\"#b22\\\">23,</ref><ref type=\\\"bibr\\\" target=\\\"#b24\\\">25,</ref><ref type=\\\"bibr\\\" target=\\\"#b25\\\">26</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"next-generation sequencing\",\n      \"label\": \"Method\",\n      \"sentence\": \"These methods include probe-based multiplexed fluorescent in situ hybridization (FISH) and next-generation sequencing (NGS)-based spatially resolved transcriptomics, each with complementary strengths in resolution, transcriptome coverage, and throughput <ref type=\\\"bibr\\\" target=\\\"#b22\\\">23,</ref><ref type=\\\"bibr\\\" target=\\\"#b24\\\">25,</ref><ref type=\\\"bibr\\\" target=\\\"#b25\\\">26</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"NGS\",\n      \"label\": \"Method\",\n      \"sentence\": \"These methods include probe-based multiplexed fluorescent in situ hybridization (FISH) and next-generation sequencing (NGS)-based spatially resolved transcriptomics, each with complementary strengths in resolution, transcriptome coverage, and throughput <ref type=\\\"bibr\\\" target=\\\"#b22\\\">23,</ref><ref type=\\\"bibr\\\" target=\\\"#b24\\\">25,</ref><ref type=\\\"bibr\\\" target=\\\"#b25\\\">26</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"striosomes\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"Although previous studies have revealed the spatial organization of striosomes within the striatal matrix and transcriptional gradients across basal ganglia nuclei in model organisms such as mice and macaques <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b26\\\">27</ref> , a comprehensive molecular and cell-type-resolved spatial atlas of the adult human basal ganglia remains lacking.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"striatal matrix\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"Although previous studies have revealed the spatial organization of striosomes within the striatal matrix and transcriptional gradients across basal ganglia nuclei in model organisms such as mice and macaques <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b26\\\">27</ref> , a comprehensive molecular and cell-type-resolved spatial atlas of the adult human basal ganglia remains lacking.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"basal ganglia nuclei\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Although previous studies have revealed the spatial organization of striosomes within the striatal matrix and transcriptional gradients across basal ganglia nuclei in model organisms such as mice and macaques <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b26\\\">27</ref> , a comprehensive molecular and cell-type-resolved spatial atlas of the adult human basal ganglia remains lacking.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"mice\",\n      \"label\": \"Species\",\n      \"sentence\": \"Although previous studies have revealed the spatial organization of striosomes within the striatal matrix and transcriptional gradients across basal ganglia nuclei in model organisms such as mice and macaques <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b26\\\">27</ref> , a comprehensive molecular and cell-type-resolved spatial atlas of the adult human basal ganglia remains lacking.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"macaques\",\n      \"label\": \"Species\",\n      \"sentence\": \"Although previous studies have revealed the spatial organization of striosomes within the striatal matrix and transcriptional gradients across basal ganglia nuclei in model organisms such as mice and macaques <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b26\\\">27</ref> , a comprehensive molecular and cell-type-resolved spatial atlas of the adult human basal ganglia remains lacking.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"Parkinson's disease\",\n      \"label\": \"Disease\",\n      \"sentence\": \"Addressing this limitation is particularly critical given the central role of the basal ganglia in human movement disorders (including Parkinson's disease and Huntington's disease), neuropsychiatric conditions, and species-specific differences in basal ganglia cellular composition and structural organization that reflect the complexity of inputs and outputs for these structures in humans <ref type=\\\"bibr\\\" target=\\\"#b6\\\">7,</ref><ref type=\\\"bibr\\\" target=\\\"#b27\\\">28</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"Huntington's disease\",\n      \"label\": \"Disease\",\n      \"sentence\": \"Addressing this limitation is particularly critical given the central role of the basal ganglia in human movement disorders (including Parkinson's disease and Huntington's disease), neuropsychiatric conditions, and species-specific differences in basal ganglia cellular composition and structural organization that reflect the complexity of inputs and outputs for these structures in humans <ref type=\\\"bibr\\\" target=\\\"#b6\\\">7,</ref><ref type=\\\"bibr\\\" target=\\\"#b27\\\">28</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"Multiplexed Error-Robust Fluorescence In Situ Hybridization\",\n      \"label\": \"Method\",\n      \"sentence\": \"Two particularly powerful approaches are Multiplexed Error-Robust Fluorescence In Situ Hybridization (MERFISH) <ref type=\\\"bibr\\\" target=\\\"#b25\\\">26,</ref><ref type=\\\"bibr\\\" target=\\\"#b28\\\">[29]</ref><ref type=\\\"bibr\\\" target=\\\"#b29\\\">[30]</ref><ref type=\\\"bibr\\\" target=\\\"#b30\\\">[31]</ref><ref type=\\\"bibr\\\" target=\\\"#b32\\\">[32]</ref><ref type=\\\"bibr\\\" target=\\\"#b33\\\">[33]</ref> and Spatio-Temporal Enhanced Resolution Omics-sequencing (Stereo-seq) <ref type=\\\"bibr\\\" target=\\\"#b34\\\">[34]</ref><ref type=\\\"bibr\\\" target=\\\"#b35\\\">[35]</ref><ref type=\\\"bibr\\\" target=\\\"#b36\\\">[36]</ref><ref type=\\\"bibr\\\" target=\\\"#b37\\\">[37]</ref><ref type=\\\"bibr\\\" target=\\\"#b38\\\">[38]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"MERFISH\",\n      \"label\": \"Method\",\n      \"sentence\": \"Two particularly powerful approaches are Multiplexed Error-Robust Fluorescence In Situ Hybridization (MERFISH) <ref type=\\\"bibr\\\" target=\\\"#b25\\\">26,</ref><ref type=\\\"bibr\\\" target=\\\"#b28\\\">[29]</ref><ref type=\\\"bibr\\\" target=\\\"#b29\\\">[30]</ref><ref type=\\\"bibr\\\" target=\\\"#b30\\\">[31]</ref><ref type=\\\"bibr\\\" target=\\\"#b32\\\">[32]</ref><ref type=\\\"bibr\\\" target=\\\"#b33\\\">[33]</ref> and Spatio-Temporal Enhanced Resolution Omics-sequencing (Stereo-seq) <ref type=\\\"bibr\\\" target=\\\"#b34\\\">[34]</ref><ref type=\\\"bibr\\\" target=\\\"#b35\\\">[35]</ref><ref type=\\\"bibr\\\" target=\\\"#b36\\\">[36]</ref><ref type=\\\"bibr\\\" target=\\\"#b37\\\">[37]</ref><ref type=\\\"bibr\\\" target=\\\"#b38\\\">[38]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"Spatio-Temporal Enhanced Resolution Omics-sequencing\",\n      \"label\": \"Method\",\n      \"sentence\": \"Two particularly powerful approaches are Multiplexed Error-Robust Fluorescence In Situ Hybridization (MERFISH) <ref type=\\\"bibr\\\" target=\\\"#b25\\\">26,</ref><ref type=\\\"bibr\\\" target=\\\"#b28\\\">[29]</ref><ref type=\\\"bibr\\\" target=\\\"#b29\\\">[30]</ref><ref type=\\\"bibr\\\" target=\\\"#b30\\\">[31]</ref><ref type=\\\"bibr\\\" target=\\\"#b32\\\">[32]</ref><ref type=\\\"bibr\\\" target=\\\"#b33\\\">[33]</ref> and Spatio-Temporal Enhanced Resolution Omics-sequencing (Stereo-seq) <ref type=\\\"bibr\\\" target=\\\"#b34\\\">[34]</ref><ref type=\\\"bibr\\\" target=\\\"#b35\\\">[35]</ref><ref type=\\\"bibr\\\" target=\\\"#b36\\\">[36]</ref><ref type=\\\"bibr\\\" target=\\\"#b37\\\">[37]</ref><ref type=\\\"bibr\\\" target=\\\"#b38\\\">[38]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"Two particularly powerful approaches are Multiplexed Error-Robust Fluorescence In Situ Hybridization (MERFISH) <ref type=\\\"bibr\\\" target=\\\"#b25\\\">26,</ref><ref type=\\\"bibr\\\" target=\\\"#b28\\\">[29]</ref><ref type=\\\"bibr\\\" target=\\\"#b29\\\">[30]</ref><ref type=\\\"bibr\\\" target=\\\"#b30\\\">[31]</ref><ref type=\\\"bibr\\\" target=\\\"#b32\\\">[32]</ref><ref type=\\\"bibr\\\" target=\\\"#b33\\\">[33]</ref> and Spatio-Temporal Enhanced Resolution Omics-sequencing (Stereo-seq) <ref type=\\\"bibr\\\" target=\\\"#b34\\\">[34]</ref><ref type=\\\"bibr\\\" target=\\\"#b35\\\">[35]</ref><ref type=\\\"bibr\\\" target=\\\"#b36\\\">[36]</ref><ref type=\\\"bibr\\\" target=\\\"#b37\\\">[37]</ref><ref type=\\\"bibr\\\" target=\\\"#b38\\\">[38]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"MERFISH\",\n      \"label\": \"Method\",\n      \"sentence\": \"MERFISH enables highly multiplexed, single-molecule RNA imaging at subcellular resolution, allowing the precise mapping of hundreds to thousands of genes within individual cells while maintaining their spatial context <ref type=\\\"bibr\\\" target=\\\"#b25\\\">26,</ref><ref type=\\\"bibr\\\" target=\\\"#b28\\\">29,</ref><ref type=\\\"bibr\\\" target=\\\"#b30\\\">31</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"single-molecule RNA imaging\",\n      \"label\": \"Method\",\n      \"sentence\": \"MERFISH enables highly multiplexed, single-molecule RNA imaging at subcellular resolution, allowing the precise mapping of hundreds to thousands of genes within individual cells while maintaining their spatial context <ref type=\\\"bibr\\\" target=\\\"#b25\\\">26,</ref><ref type=\\\"bibr\\\" target=\\\"#b28\\\">29,</ref><ref type=\\\"bibr\\\" target=\\\"#b30\\\">31</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"mouse\",\n      \"label\": \"Species\",\n      \"sentence\": \"This approach has been instrumental in decoding cell-type organization across cortical and subcortical structures in multiple species that reveal both conserved and species-specific architectural principles in the mouse, monkey and human brains <ref type=\\\"bibr\\\" target=\\\"#b13\\\">14,</ref><ref type=\\\"bibr\\\" target=\\\"#b26\\\">27,</ref><ref type=\\\"bibr\\\" target=\\\"#b36\\\">36,</ref><ref type=\\\"bibr\\\" target=\\\"#b39\\\">[39]</ref><ref type=\\\"bibr\\\" target=\\\"#b40\\\">[40]</ref><ref type=\\\"bibr\\\" target=\\\"#b41\\\">[41]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"monkey\",\n      \"label\": \"Species\",\n      \"sentence\": \"This approach has been instrumental in decoding cell-type organization across cortical and subcortical structures in multiple species that reveal both conserved and species-specific architectural principles in the mouse, monkey and human brains <ref type=\\\"bibr\\\" target=\\\"#b13\\\">14,</ref><ref type=\\\"bibr\\\" target=\\\"#b26\\\">27,</ref><ref type=\\\"bibr\\\" target=\\\"#b36\\\">36,</ref><ref type=\\\"bibr\\\" target=\\\"#b39\\\">[39]</ref><ref type=\\\"bibr\\\" target=\\\"#b40\\\">[40]</ref><ref type=\\\"bibr\\\" target=\\\"#b41\\\">[41]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"human brains\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"This approach has been instrumental in decoding cell-type organization across cortical and subcortical structures in multiple species that reveal both conserved and species-specific architectural principles in the mouse, monkey and human brains <ref type=\\\"bibr\\\" target=\\\"#b13\\\">14,</ref><ref type=\\\"bibr\\\" target=\\\"#b26\\\">27,</ref><ref type=\\\"bibr\\\" target=\\\"#b36\\\">36,</ref><ref type=\\\"bibr\\\" target=\\\"#b39\\\">[39]</ref><ref type=\\\"bibr\\\" target=\\\"#b40\\\">[40]</ref><ref type=\\\"bibr\\\" target=\\\"#b41\\\">[41]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"MERFISH\",\n      \"label\": \"Method\",\n      \"sentence\": \"Subsequent advancements in MERFISH have expanded high-resolution spatial transcriptomics into a multimodal platform capable of imaging DNA, RNA, and proteins within the same tissue specimen <ref type=\\\"bibr\\\" target=\\\"#b42\\\">42,</ref><ref type=\\\"bibr\\\" target=\\\"#b43\\\">43</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"DNA\",\n      \"label\": \"Chemical\",\n      \"sentence\": \"Subsequent advancements in MERFISH have expanded high-resolution spatial transcriptomics into a multimodal platform capable of imaging DNA, RNA, and proteins within the same tissue specimen <ref type=\\\"bibr\\\" target=\\\"#b42\\\">42,</ref><ref type=\\\"bibr\\\" target=\\\"#b43\\\">43</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"RNA\",\n      \"label\": \"Chemical\",\n      \"sentence\": \"Subsequent advancements in MERFISH have expanded high-resolution spatial transcriptomics into a multimodal platform capable of imaging DNA, RNA, and proteins within the same tissue specimen <ref type=\\\"bibr\\\" target=\\\"#b42\\\">42,</ref><ref type=\\\"bibr\\\" target=\\\"#b43\\\">43</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"proteins\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Subsequent advancements in MERFISH have expanded high-resolution spatial transcriptomics into a multimodal platform capable of imaging DNA, RNA, and proteins within the same tissue specimen <ref type=\\\"bibr\\\" target=\\\"#b42\\\">42,</ref><ref type=\\\"bibr\\\" target=\\\"#b43\\\">43</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"Most recently, the incorporation of modified probe designs, high-throughput microscopy, and microfluidic automation has led to MERFISH+ <ref type=\\\"bibr\\\" target=\\\"#b44\\\">44</ref> , an enhanced implementation that enables imaging across multi-centimeter-scale tissue areas.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"high-throughput microscopy\",\n      \"label\": \"Method\",\n      \"sentence\": \"Most recently, the incorporation of modified probe designs, high-throughput microscopy, and microfluidic automation has led to MERFISH+ <ref type=\\\"bibr\\\" target=\\\"#b44\\\">44</ref> , an enhanced implementation that enables imaging across multi-centimeter-scale tissue areas.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"microfluidic automation\",\n      \"label\": \"Method\",\n      \"sentence\": \"Most recently, the incorporation of modified probe designs, high-throughput microscopy, and microfluidic automation has led to MERFISH+ <ref type=\\\"bibr\\\" target=\\\"#b44\\\">44</ref> , an enhanced implementation that enables imaging across multi-centimeter-scale tissue areas.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"Stereo-seq employs DNA nanoball-patterned arrays to generate spatially barcoded cDNA libraries that capture transcriptomes from large tissue sections at high resolution <ref type=\\\"bibr\\\" target=\\\"#b34\\\">34,</ref><ref type=\\\"bibr\\\" target=\\\"#b36\\\">[36]</ref><ref type=\\\"bibr\\\" target=\\\"#b37\\\">[37]</ref><ref type=\\\"bibr\\\" target=\\\"#b38\\\">[38]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"DNA nanoball-patterned arrays\",\n      \"label\": \"Method\",\n      \"sentence\": \"Stereo-seq employs DNA nanoball-patterned arrays to generate spatially barcoded cDNA libraries that capture transcriptomes from large tissue sections at high resolution <ref type=\\\"bibr\\\" target=\\\"#b34\\\">34,</ref><ref type=\\\"bibr\\\" target=\\\"#b36\\\">[36]</ref><ref type=\\\"bibr\\\" target=\\\"#b37\\\">[37]</ref><ref type=\\\"bibr\\\" target=\\\"#b38\\\">[38]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"cDNA libraries\",\n      \"label\": \"NucleicAcid\",\n      \"sentence\": \"Stereo-seq employs DNA nanoball-patterned arrays to generate spatially barcoded cDNA libraries that capture transcriptomes from large tissue sections at high resolution <ref type=\\\"bibr\\\" target=\\\"#b34\\\">34,</ref><ref type=\\\"bibr\\\" target=\\\"#b36\\\">[36]</ref><ref type=\\\"bibr\\\" target=\\\"#b37\\\">[37]</ref><ref type=\\\"bibr\\\" target=\\\"#b38\\\">[38]</ref> .\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"In this study, we harness complementary strengths of MERFISH+ and Stereo-seq to generate a multiscale spatial transcriptomic atlas of the human basal ganglia to capture spatial gene-expression patterns, specific cell types, and molecular architectures across multi-centimeter tissue sections from four neurotypical adult postmortem donors.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"In this study, we harness complementary strengths of MERFISH+ and Stereo-seq to generate a multiscale spatial transcriptomic atlas of the human basal ganglia to capture spatial gene-expression patterns, specific cell types, and molecular architectures across multi-centimeter tissue sections from four neurotypical adult postmortem donors.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"human basal ganglia\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"In this study, we harness complementary strengths of MERFISH+ and Stereo-seq to generate a multiscale spatial transcriptomic atlas of the human basal ganglia to capture spatial gene-expression patterns, specific cell types, and molecular architectures across multi-centimeter tissue sections from four neurotypical adult postmortem donors.\",\n      \"paper_location\": \"INTRODUCTION\"\n    },\n    {\n      \"entity\": \"human brain\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"We obtained four postmortem human brain fresh frozen slabs across anterior-toposterior segments of the basal ganglia encompassing the caudate nucleus, putamen, nucleus accumbens (NAc), globus pallidus, and internal capsule (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1A</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"basal ganglia\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"We obtained four postmortem human brain fresh frozen slabs across anterior-toposterior segments of the basal ganglia encompassing the caudate nucleus, putamen, nucleus accumbens (NAc), globus pallidus, and internal capsule (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1A</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"caudate nucleus\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"We obtained four postmortem human brain fresh frozen slabs across anterior-toposterior segments of the basal ganglia encompassing the caudate nucleus, putamen, nucleus accumbens (NAc), globus pallidus, and internal capsule (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1A</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"putamen\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"We obtained four postmortem human brain fresh frozen slabs across anterior-toposterior segments of the basal ganglia encompassing the caudate nucleus, putamen, nucleus accumbens (NAc), globus pallidus, and internal capsule (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1A</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"nucleus accumbens\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"We obtained four postmortem human brain fresh frozen slabs across anterior-toposterior segments of the basal ganglia encompassing the caudate nucleus, putamen, nucleus accumbens (NAc), globus pallidus, and internal capsule (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1A</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"NAc\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"We obtained four postmortem human brain fresh frozen slabs across anterior-toposterior segments of the basal ganglia encompassing the caudate nucleus, putamen, nucleus accumbens (NAc), globus pallidus, and internal capsule (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1A</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"globus pallidus\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"We obtained four postmortem human brain fresh frozen slabs across anterior-toposterior segments of the basal ganglia encompassing the caudate nucleus, putamen, nucleus accumbens (NAc), globus pallidus, and internal capsule (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1A</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"internal capsule\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"We obtained four postmortem human brain fresh frozen slabs across anterior-toposterior segments of the basal ganglia encompassing the caudate nucleus, putamen, nucleus accumbens (NAc), globus pallidus, and internal capsule (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1A</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"neurodegenerative disease\",\n      \"label\": \"Disease\",\n      \"sentence\": \"Tissue samples were collected from four neurotypical Caucasian male donors, aged 36-53 years, with no pathological evidence of neurodegenerative disease or other brain injury.\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"brain injury\",\n      \"label\": \"Disease\",\n      \"sentence\": \"Tissue samples were collected from four neurotypical Caucasian male donors, aged 36-53 years, with no pathological evidence of neurodegenerative disease or other brain injury.\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"RNA Integrity Number\",\n      \"label\": \"Measurement\",\n      \"sentence\": \"RNA quality was confirmed by RNA Integrity Number (RIN) values ranging from 7.0 to 8.4, indicating excellent preservation for downstream analyses (Table <ref type=\\\"table\\\">S1</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"RIN\",\n      \"label\": \"Measurement\",\n      \"sentence\": \"RNA quality was confirmed by RNA Integrity Number (RIN) values ranging from 7.0 to 8.4, indicating excellent preservation for downstream analyses (Table <ref type=\\\"table\\\">S1</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"For spatial transcriptomic profiling, 10 µm-thick cryostat sections were processed for Stereo-seq with the coverage area size of up to 20 mm × 30 mm, while adjacent 16 µmthick sections were selected for MERFISH+ imaging with an area size of up to 43 mm × 30 mm (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1A</ref>, Supplementary Movie 1, Table <ref type=\\\"table\\\">1S</ref> & Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1S</ref> in Supplemental Information).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"For spatial transcriptomic profiling, 10 µm-thick cryostat sections were processed for Stereo-seq with the coverage area size of up to 20 mm × 30 mm, while adjacent 16 µmthick sections were selected for MERFISH+ imaging with an area size of up to 43 mm × 30 mm (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1A</ref>, Supplementary Movie 1, Table <ref type=\\\"table\\\">1S</ref> & Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1S</ref> in Supplemental Information).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"A single 20mm x 30mm Stereo-seq chip captured ~0.9 million cells (with ~0.29 million cells selected for cell type classification), which were clustered into 10 different major cell types (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1B</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"Across four Stereo-seq assays, we obtained genome-wide transcriptomic coverage of ~70,000 genes across ~3.2 million cells.\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"In parallel, comparable numbers of cells per area were captured by MERFISH+ on adjacent sections targeting a panel of 673 selected genes (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1C</ref>, Supplemental Data).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"Four MERFISH+ experiments captured ~3.9 million cells at single-molecule resolution over multi-centimeter-scale tissue areas.\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"Although Stereo-seq detects fewer transcripts per cell due to its sequencing-based capture efficiency, the expression levels of the 673 MERFISH+ target genes show strong cross-platform correlation between the two modalities (Pearson correlation coefficient of 0.71), demonstrating the quantitative agreement of spatial gene-expression measurements across methodological approaches (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">S1B</ref>) with high reproducibility across and between experiments (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">S1C-D</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"Although Stereo-seq detects fewer transcripts per cell due to its sequencing-based capture efficiency, the expression levels of the 673 MERFISH+ target genes show strong cross-platform correlation between the two modalities (Pearson correlation coefficient of 0.71), demonstrating the quantitative agreement of spatial gene-expression measurements across methodological approaches (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">S1B</ref>) with high reproducibility across and between experiments (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">S1C-D</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"Pearson correlation coefficient\",\n      \"label\": \"Measurement\",\n      \"sentence\": \"Although Stereo-seq detects fewer transcripts per cell due to its sequencing-based capture efficiency, the expression levels of the 673 MERFISH+ target genes show strong cross-platform correlation between the two modalities (Pearson correlation coefficient of 0.71), demonstrating the quantitative agreement of spatial gene-expression measurements across methodological approaches (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">S1B</ref>) with high reproducibility across and between experiments (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">S1C-D</ref>).\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"Given these respective advantages and limitations, we used MERFISH+ to focus on fine-scale cell definition and subcellular localization, and Stereoseq for exploring large scale gene expression patterns across centimeter scale tissue.\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"Stereoseq\",\n      \"label\": \"Method\",\n      \"sentence\": \"Given these respective advantages and limitations, we used MERFISH+ to focus on fine-scale cell definition and subcellular localization, and Stereoseq for exploring large scale gene expression patterns across centimeter scale tissue.\",\n      \"paper_location\": \"Multi-region, subcellular-resolution spatial transcriptomic mapping of human basal ganglia\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"Using MERFISH+, we resolve approximately 60 transcriptionally distinct cell types that are separable when embedded in Uniform Manifold Approximation and Projection (UMAP) space of gene-expression data (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2A</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"Uniform Manifold Approximation and Projection\",\n      \"label\": \"Method\",\n      \"sentence\": \"Using MERFISH+, we resolve approximately 60 transcriptionally distinct cell types that are separable when embedded in Uniform Manifold Approximation and Projection (UMAP) space of gene-expression data (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2A</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"UMAP\",\n      \"label\": \"Method\",\n      \"sentence\": \"Using MERFISH+, we resolve approximately 60 transcriptionally distinct cell types that are separable when embedded in Uniform Manifold Approximation and Projection (UMAP) space of gene-expression data (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2A</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"single-nucleus RNA sequencing\",\n      \"label\": \"Method\",\n      \"sentence\": \"These cell types show expression of expected corresponding marker genes (Figure <ref type=\\\"figure\\\" target=\\\"#fig_9\\\">S2</ref>) and demonstrate strong correspondence with single-nucleus RNA sequencing (snRNA-seq) datasets from the Allen Institute Mammalian Basal Ganglia Consensus Cell Type Atlas (Allen Institute; RRID:SCR_024672) <ref type=\\\"bibr\\\" target=\\\"#b45\\\">45</ref> (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2B</ref>), validating our cell type classification.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"snRNA-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"These cell types show expression of expected corresponding marker genes (Figure <ref type=\\\"figure\\\" target=\\\"#fig_9\\\">S2</ref>) and demonstrate strong correspondence with single-nucleus RNA sequencing (snRNA-seq) datasets from the Allen Institute Mammalian Basal Ganglia Consensus Cell Type Atlas (Allen Institute; RRID:SCR_024672) <ref type=\\\"bibr\\\" target=\\\"#b45\\\">45</ref> (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2B</ref>), validating our cell type classification.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"Allen Institute Mammalian Basal Ganglia Consensus Cell Type Atlas\",\n      \"label\": \"DataResource\",\n      \"sentence\": \"These cell types show expression of expected corresponding marker genes (Figure <ref type=\\\"figure\\\" target=\\\"#fig_9\\\">S2</ref>) and demonstrate strong correspondence with single-nucleus RNA sequencing (snRNA-seq) datasets from the Allen Institute Mammalian Basal Ganglia Consensus Cell Type Atlas (Allen Institute; RRID:SCR_024672) <ref type=\\\"bibr\\\" target=\\\"#b45\\\">45</ref> (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2B</ref>), validating our cell type classification.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"caudal and lateral ganglionic eminence\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Among these are eight neuronal classes that include caudal and lateral ganglionic eminence (CGE/LGE)-derived GABAergic neurons, which are dominated by medium spiny neurons (MSNs) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3A</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"CGE/LGE\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Among these are eight neuronal classes that include caudal and lateral ganglionic eminence (CGE/LGE)-derived GABAergic neurons, which are dominated by medium spiny neurons (MSNs) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3A</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GABAergic neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Among these are eight neuronal classes that include caudal and lateral ganglionic eminence (CGE/LGE)-derived GABAergic neurons, which are dominated by medium spiny neurons (MSNs) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3A</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"medium spiny neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Among these are eight neuronal classes that include caudal and lateral ganglionic eminence (CGE/LGE)-derived GABAergic neurons, which are dominated by medium spiny neurons (MSNs) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3A</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Among these are eight neuronal classes that include caudal and lateral ganglionic eminence (CGE/LGE)-derived GABAergic neurons, which are dominated by medium spiny neurons (MSNs) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3A</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"We further examined MSNs -the predominant neuronal population of the striatum that are essential for modulating movement control in the basal ganglia circuit <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3</ref> .\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"striatum\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"We further examined MSNs -the predominant neuronal population of the striatum that are essential for modulating movement control in the basal ganglia circuit <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3</ref> .\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"basal ganglia circuit\",\n      \"label\": \"NeuralCircuit\",\n      \"sentence\": \"We further examined MSNs -the predominant neuronal population of the striatum that are essential for modulating movement control in the basal ganglia circuit <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3</ref> .\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"MSNs segregate into two canonical classes -D1 and D2 MSNs -further subdivided into six major dopaminergic MSN subtypes: STRd D1 Striosome, STRd D1 Matrix, STRv D1, STRd D2 Striosome, STRd D2 Matrix and STRv D2 (dorsal striatum, STRd; ventral striatum, STRv) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2D-E</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3B</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"D1\",\n      \"label\": \"CellType\",\n      \"sentence\": \"MSNs segregate into two canonical classes -D1 and D2 MSNs -further subdivided into six major dopaminergic MSN subtypes: STRd D1 Striosome, STRd D1 Matrix, STRv D1, STRd D2 Striosome, STRd D2 Matrix and STRv D2 (dorsal striatum, STRd; ventral striatum, STRv) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2D-E</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3B</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"D2 MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"MSNs segregate into two canonical classes -D1 and D2 MSNs -further subdivided into six major dopaminergic MSN subtypes: STRd D1 Striosome, STRd D1 Matrix, STRv D1, STRd D2 Striosome, STRd D2 Matrix and STRv D2 (dorsal striatum, STRd; ventral striatum, STRv) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2D-E</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3B</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRd D1 Striosome\",\n      \"label\": \"CellType\",\n      \"sentence\": \"MSNs segregate into two canonical classes -D1 and D2 MSNs -further subdivided into six major dopaminergic MSN subtypes: STRd D1 Striosome, STRd D1 Matrix, STRv D1, STRd D2 Striosome, STRd D2 Matrix and STRv D2 (dorsal striatum, STRd; ventral striatum, STRv) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2D-E</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3B</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRd D1 Matrix\",\n      \"label\": \"CellType\",\n      \"sentence\": \"MSNs segregate into two canonical classes -D1 and D2 MSNs -further subdivided into six major dopaminergic MSN subtypes: STRd D1 Striosome, STRd D1 Matrix, STRv D1, STRd D2 Striosome, STRd D2 Matrix and STRv D2 (dorsal striatum, STRd; ventral striatum, STRv) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2D-E</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3B</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRv D1\",\n      \"label\": \"CellType\",\n      \"sentence\": \"MSNs segregate into two canonical classes -D1 and D2 MSNs -further subdivided into six major dopaminergic MSN subtypes: STRd D1 Striosome, STRd D1 Matrix, STRv D1, STRd D2 Striosome, STRd D2 Matrix and STRv D2 (dorsal striatum, STRd; ventral striatum, STRv) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2D-E</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3B</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRd D2 Striosome\",\n      \"label\": \"CellType\",\n      \"sentence\": \"MSNs segregate into two canonical classes -D1 and D2 MSNs -further subdivided into six major dopaminergic MSN subtypes: STRd D1 Striosome, STRd D1 Matrix, STRv D1, STRd D2 Striosome, STRd D2 Matrix and STRv D2 (dorsal striatum, STRd; ventral striatum, STRv) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2D-E</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3B</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRd D2 Matrix\",\n      \"label\": \"CellType\",\n      \"sentence\": \"MSNs segregate into two canonical classes -D1 and D2 MSNs -further subdivided into six major dopaminergic MSN subtypes: STRd D1 Striosome, STRd D1 Matrix, STRv D1, STRd D2 Striosome, STRd D2 Matrix and STRv D2 (dorsal striatum, STRd; ventral striatum, STRv) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2D-E</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3B</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRv D2\",\n      \"label\": \"CellType\",\n      \"sentence\": \"MSNs segregate into two canonical classes -D1 and D2 MSNs -further subdivided into six major dopaminergic MSN subtypes: STRd D1 Striosome, STRd D1 Matrix, STRv D1, STRd D2 Striosome, STRd D2 Matrix and STRv D2 (dorsal striatum, STRd; ventral striatum, STRv) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2D-E</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3B</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"dorsal striatum\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"MSNs segregate into two canonical classes -D1 and D2 MSNs -further subdivided into six major dopaminergic MSN subtypes: STRd D1 Striosome, STRd D1 Matrix, STRv D1, STRd D2 Striosome, STRd D2 Matrix and STRv D2 (dorsal striatum, STRd; ventral striatum, STRv) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2D-E</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3B</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRd\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"MSNs segregate into two canonical classes -D1 and D2 MSNs -further subdivided into six major dopaminergic MSN subtypes: STRd D1 Striosome, STRd D1 Matrix, STRv D1, STRd D2 Striosome, STRd D2 Matrix and STRv D2 (dorsal striatum, STRd; ventral striatum, STRv) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2D-E</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3B</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"ventral striatum\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"MSNs segregate into two canonical classes -D1 and D2 MSNs -further subdivided into six major dopaminergic MSN subtypes: STRd D1 Striosome, STRd D1 Matrix, STRv D1, STRd D2 Striosome, STRd D2 Matrix and STRv D2 (dorsal striatum, STRd; ventral striatum, STRv) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2D-E</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3B</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRv\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"MSNs segregate into two canonical classes -D1 and D2 MSNs -further subdivided into six major dopaminergic MSN subtypes: STRd D1 Striosome, STRd D1 Matrix, STRv D1, STRd D2 Striosome, STRd D2 Matrix and STRv D2 (dorsal striatum, STRd; ventral striatum, STRv) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2D-E</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_10\\\">S3B</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"D1-D2 hybrid\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The D1-D2 hybrid (STR D1D2 Hybrid), striosome-matrix hybrid (STRd D2 Hybrid), and D1 NUDAP (Neurochemically Unique Domains in the Accumbens and Putamen) (STRv D1 NUDAP) MSNs are primarily localized within the putamen, caudate nucleus, and nucleus accumbens (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2G</ref>), but are largely absent from the globus pallidus and internal capsule.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STR D1D2 Hybrid\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The D1-D2 hybrid (STR D1D2 Hybrid), striosome-matrix hybrid (STRd D2 Hybrid), and D1 NUDAP (Neurochemically Unique Domains in the Accumbens and Putamen) (STRv D1 NUDAP) MSNs are primarily localized within the putamen, caudate nucleus, and nucleus accumbens (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2G</ref>), but are largely absent from the globus pallidus and internal capsule.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"striosome-matrix hybrid\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The D1-D2 hybrid (STR D1D2 Hybrid), striosome-matrix hybrid (STRd D2 Hybrid), and D1 NUDAP (Neurochemically Unique Domains in the Accumbens and Putamen) (STRv D1 NUDAP) MSNs are primarily localized within the putamen, caudate nucleus, and nucleus accumbens (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2G</ref>), but are largely absent from the globus pallidus and internal capsule.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRd D2 Hybrid\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The D1-D2 hybrid (STR D1D2 Hybrid), striosome-matrix hybrid (STRd D2 Hybrid), and D1 NUDAP (Neurochemically Unique Domains in the Accumbens and Putamen) (STRv D1 NUDAP) MSNs are primarily localized within the putamen, caudate nucleus, and nucleus accumbens (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2G</ref>), but are largely absent from the globus pallidus and internal capsule.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"Neurochemically Unique Domains in the Accumbens and Putamen\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"The D1-D2 hybrid (STR D1D2 Hybrid), striosome-matrix hybrid (STRd D2 Hybrid), and D1 NUDAP (Neurochemically Unique Domains in the Accumbens and Putamen) (STRv D1 NUDAP) MSNs are primarily localized within the putamen, caudate nucleus, and nucleus accumbens (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2G</ref>), but are largely absent from the globus pallidus and internal capsule.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRv D1 NUDAP\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The D1-D2 hybrid (STR D1D2 Hybrid), striosome-matrix hybrid (STRd D2 Hybrid), and D1 NUDAP (Neurochemically Unique Domains in the Accumbens and Putamen) (STRv D1 NUDAP) MSNs are primarily localized within the putamen, caudate nucleus, and nucleus accumbens (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2G</ref>), but are largely absent from the globus pallidus and internal capsule.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"DRD1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"The mixed state of these cell populations is corroborated by colocalization within individual neurons of multiple transcripts of DRD1, DRD2, and prodynorphin (PDYN), a canonical striosomal MSN marker (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2H</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"DRD2\",\n      \"label\": \"Gene\",\n      \"sentence\": \"The mixed state of these cell populations is corroborated by colocalization within individual neurons of multiple transcripts of DRD1, DRD2, and prodynorphin (PDYN), a canonical striosomal MSN marker (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2H</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"prodynorphin\",\n      \"label\": \"Neuropeptide\",\n      \"sentence\": \"The mixed state of these cell populations is corroborated by colocalization within individual neurons of multiple transcripts of DRD1, DRD2, and prodynorphin (PDYN), a canonical striosomal MSN marker (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2H</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"PDYN\",\n      \"label\": \"Gene\",\n      \"sentence\": \"The mixed state of these cell populations is corroborated by colocalization within individual neurons of multiple transcripts of DRD1, DRD2, and prodynorphin (PDYN), a canonical striosomal MSN marker (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2H</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GPi Shell\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Analysis of two posterior sections encompassing both internal (GPi) and external (GPe) segments reveal six transcriptionally distinct GP-associated neuronal classes: GPi Shell, GPi Core, GPin-BF Cholinergic, CN LHX6-GBX1, GPe SOX6-CTXND1, and GPe MEIS2-SOX6 (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2I</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">S4</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GPi Core\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Analysis of two posterior sections encompassing both internal (GPi) and external (GPe) segments reveal six transcriptionally distinct GP-associated neuronal classes: GPi Shell, GPi Core, GPin-BF Cholinergic, CN LHX6-GBX1, GPe SOX6-CTXND1, and GPe MEIS2-SOX6 (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2I</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">S4</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GPin-BF Cholinergic\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Analysis of two posterior sections encompassing both internal (GPi) and external (GPe) segments reveal six transcriptionally distinct GP-associated neuronal classes: GPi Shell, GPi Core, GPin-BF Cholinergic, CN LHX6-GBX1, GPe SOX6-CTXND1, and GPe MEIS2-SOX6 (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2I</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">S4</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"CN LHX6-GBX1\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Analysis of two posterior sections encompassing both internal (GPi) and external (GPe) segments reveal six transcriptionally distinct GP-associated neuronal classes: GPi Shell, GPi Core, GPin-BF Cholinergic, CN LHX6-GBX1, GPe SOX6-CTXND1, and GPe MEIS2-SOX6 (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2I</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">S4</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GPe SOX6-CTXND1\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Analysis of two posterior sections encompassing both internal (GPi) and external (GPe) segments reveal six transcriptionally distinct GP-associated neuronal classes: GPi Shell, GPi Core, GPin-BF Cholinergic, CN LHX6-GBX1, GPe SOX6-CTXND1, and GPe MEIS2-SOX6 (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2I</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">S4</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GPe MEIS2-SOX6\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Analysis of two posterior sections encompassing both internal (GPi) and external (GPe) segments reveal six transcriptionally distinct GP-associated neuronal classes: GPi Shell, GPi Core, GPin-BF Cholinergic, CN LHX6-GBX1, GPe SOX6-CTXND1, and GPe MEIS2-SOX6 (Figure <ref type=\\\"figure\\\" target=\\\"#fig_3\\\">2I</ref>, Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">S4</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"immune\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Four major categories are observed: immune, vascular, astrocytic/ependymal (Astro-Epen), and oligodendrocyte lineage (OPC-Oligo) cells, each exhibiting characteristic regional patterns.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"vascular\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Four major categories are observed: immune, vascular, astrocytic/ependymal (Astro-Epen), and oligodendrocyte lineage (OPC-Oligo) cells, each exhibiting characteristic regional patterns.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"astrocytic/ependymal\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Four major categories are observed: immune, vascular, astrocytic/ependymal (Astro-Epen), and oligodendrocyte lineage (OPC-Oligo) cells, each exhibiting characteristic regional patterns.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"Astro-Epen\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Four major categories are observed: immune, vascular, astrocytic/ependymal (Astro-Epen), and oligodendrocyte lineage (OPC-Oligo) cells, each exhibiting characteristic regional patterns.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"oligodendrocyte lineage\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Four major categories are observed: immune, vascular, astrocytic/ependymal (Astro-Epen), and oligodendrocyte lineage (OPC-Oligo) cells, each exhibiting characteristic regional patterns.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"OPC-Oligo\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Four major categories are observed: immune, vascular, astrocytic/ependymal (Astro-Epen), and oligodendrocyte lineage (OPC-Oligo) cells, each exhibiting characteristic regional patterns.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"committed oligodendrocyte progenitors\",\n      \"label\": \"CellType\",\n      \"sentence\": \"This group includes five transcriptionally distinct subtypes: committed oligodendrocyte progenitors (COP), oligodendrocyte progenitor cells (OPC), Oligo PLEKHG1, immune-associated oligodendrocytes (ImOligo), and Oligo OPALIN (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3A</ref>-right).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"COP\",\n      \"label\": \"CellType\",\n      \"sentence\": \"This group includes five transcriptionally distinct subtypes: committed oligodendrocyte progenitors (COP), oligodendrocyte progenitor cells (OPC), Oligo PLEKHG1, immune-associated oligodendrocytes (ImOligo), and Oligo OPALIN (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3A</ref>-right).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"oligodendrocyte progenitor cells\",\n      \"label\": \"CellType\",\n      \"sentence\": \"This group includes five transcriptionally distinct subtypes: committed oligodendrocyte progenitors (COP), oligodendrocyte progenitor cells (OPC), Oligo PLEKHG1, immune-associated oligodendrocytes (ImOligo), and Oligo OPALIN (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3A</ref>-right).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"OPC\",\n      \"label\": \"CellType\",\n      \"sentence\": \"This group includes five transcriptionally distinct subtypes: committed oligodendrocyte progenitors (COP), oligodendrocyte progenitor cells (OPC), Oligo PLEKHG1, immune-associated oligodendrocytes (ImOligo), and Oligo OPALIN (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3A</ref>-right).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"Oligo PLEKHG1\",\n      \"label\": \"CellType\",\n      \"sentence\": \"This group includes five transcriptionally distinct subtypes: committed oligodendrocyte progenitors (COP), oligodendrocyte progenitor cells (OPC), Oligo PLEKHG1, immune-associated oligodendrocytes (ImOligo), and Oligo OPALIN (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3A</ref>-right).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"immune-associated oligodendrocytes\",\n      \"label\": \"CellType\",\n      \"sentence\": \"This group includes five transcriptionally distinct subtypes: committed oligodendrocyte progenitors (COP), oligodendrocyte progenitor cells (OPC), Oligo PLEKHG1, immune-associated oligodendrocytes (ImOligo), and Oligo OPALIN (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3A</ref>-right).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"ImOligo\",\n      \"label\": \"CellType\",\n      \"sentence\": \"This group includes five transcriptionally distinct subtypes: committed oligodendrocyte progenitors (COP), oligodendrocyte progenitor cells (OPC), Oligo PLEKHG1, immune-associated oligodendrocytes (ImOligo), and Oligo OPALIN (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3A</ref>-right).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"Oligo OPALIN\",\n      \"label\": \"CellType\",\n      \"sentence\": \"This group includes five transcriptionally distinct subtypes: committed oligodendrocyte progenitors (COP), oligodendrocyte progenitor cells (OPC), Oligo PLEKHG1, immune-associated oligodendrocytes (ImOligo), and Oligo OPALIN (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3A</ref>-right).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"astrocytes\",\n      \"label\": \"CellType\",\n      \"sentence\": \"We found that astrocytes also exhibit pronounced regional heterogeneity.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"astrocyte subtypes\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Subclustering analysis identifies six astrocyte subtypes, each defined by distinct marker gene combinations that include aquaporin 4 (AQP4), aquaporin 1 (AQP1), solute carrier family 6 member 11 (SLC6A11), chitinase 3-like 1 (CHI3L1), and adhesion G protein-coupled receptor V1 (ADGRV1) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3D</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"aquaporin 4\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Subclustering analysis identifies six astrocyte subtypes, each defined by distinct marker gene combinations that include aquaporin 4 (AQP4), aquaporin 1 (AQP1), solute carrier family 6 member 11 (SLC6A11), chitinase 3-like 1 (CHI3L1), and adhesion G protein-coupled receptor V1 (ADGRV1) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3D</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"AQP4\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Subclustering analysis identifies six astrocyte subtypes, each defined by distinct marker gene combinations that include aquaporin 4 (AQP4), aquaporin 1 (AQP1), solute carrier family 6 member 11 (SLC6A11), chitinase 3-like 1 (CHI3L1), and adhesion G protein-coupled receptor V1 (ADGRV1) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3D</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"aquaporin 1\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Subclustering analysis identifies six astrocyte subtypes, each defined by distinct marker gene combinations that include aquaporin 4 (AQP4), aquaporin 1 (AQP1), solute carrier family 6 member 11 (SLC6A11), chitinase 3-like 1 (CHI3L1), and adhesion G protein-coupled receptor V1 (ADGRV1) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3D</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"AQP1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Subclustering analysis identifies six astrocyte subtypes, each defined by distinct marker gene combinations that include aquaporin 4 (AQP4), aquaporin 1 (AQP1), solute carrier family 6 member 11 (SLC6A11), chitinase 3-like 1 (CHI3L1), and adhesion G protein-coupled receptor V1 (ADGRV1) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3D</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"solute carrier family 6 member 11\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Subclustering analysis identifies six astrocyte subtypes, each defined by distinct marker gene combinations that include aquaporin 4 (AQP4), aquaporin 1 (AQP1), solute carrier family 6 member 11 (SLC6A11), chitinase 3-like 1 (CHI3L1), and adhesion G protein-coupled receptor V1 (ADGRV1) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3D</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"SLC6A11\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Subclustering analysis identifies six astrocyte subtypes, each defined by distinct marker gene combinations that include aquaporin 4 (AQP4), aquaporin 1 (AQP1), solute carrier family 6 member 11 (SLC6A11), chitinase 3-like 1 (CHI3L1), and adhesion G protein-coupled receptor V1 (ADGRV1) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3D</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"chitinase 3-like 1\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Subclustering analysis identifies six astrocyte subtypes, each defined by distinct marker gene combinations that include aquaporin 4 (AQP4), aquaporin 1 (AQP1), solute carrier family 6 member 11 (SLC6A11), chitinase 3-like 1 (CHI3L1), and adhesion G protein-coupled receptor V1 (ADGRV1) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3D</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"CHI3L1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Subclustering analysis identifies six astrocyte subtypes, each defined by distinct marker gene combinations that include aquaporin 4 (AQP4), aquaporin 1 (AQP1), solute carrier family 6 member 11 (SLC6A11), chitinase 3-like 1 (CHI3L1), and adhesion G protein-coupled receptor V1 (ADGRV1) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3D</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"adhesion G protein-coupled receptor V1\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Subclustering analysis identifies six astrocyte subtypes, each defined by distinct marker gene combinations that include aquaporin 4 (AQP4), aquaporin 1 (AQP1), solute carrier family 6 member 11 (SLC6A11), chitinase 3-like 1 (CHI3L1), and adhesion G protein-coupled receptor V1 (ADGRV1) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3D</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"ADGRV1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Subclustering analysis identifies six astrocyte subtypes, each defined by distinct marker gene combinations that include aquaporin 4 (AQP4), aquaporin 1 (AQP1), solute carrier family 6 member 11 (SLC6A11), chitinase 3-like 1 (CHI3L1), and adhesion G protein-coupled receptor V1 (ADGRV1) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3D</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"WM_Astro\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Spatial mapping reveal distinct topographic specializations (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3G</ref> and Figure <ref type=\\\"figure\\\" target=\\\"#fig_11\\\">S6B-C</ref>): WM_Astro cells, characterized by high AQP1 expression, are localized primarily to white matter tracts and the peripheral globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3H</ref>); STR_Astro-2 cells, defined by ADGRV1, are enriched in striatum (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3I</ref>), and GP-TH_Astro cells, expressing SLC6A11, are selectively concentrated in the globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3J</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"AQP1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Spatial mapping reveal distinct topographic specializations (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3G</ref> and Figure <ref type=\\\"figure\\\" target=\\\"#fig_11\\\">S6B-C</ref>): WM_Astro cells, characterized by high AQP1 expression, are localized primarily to white matter tracts and the peripheral globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3H</ref>); STR_Astro-2 cells, defined by ADGRV1, are enriched in striatum (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3I</ref>), and GP-TH_Astro cells, expressing SLC6A11, are selectively concentrated in the globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3J</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"white matter tracts\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Spatial mapping reveal distinct topographic specializations (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3G</ref> and Figure <ref type=\\\"figure\\\" target=\\\"#fig_11\\\">S6B-C</ref>): WM_Astro cells, characterized by high AQP1 expression, are localized primarily to white matter tracts and the peripheral globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3H</ref>); STR_Astro-2 cells, defined by ADGRV1, are enriched in striatum (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3I</ref>), and GP-TH_Astro cells, expressing SLC6A11, are selectively concentrated in the globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3J</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STR_Astro-2\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Spatial mapping reveal distinct topographic specializations (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3G</ref> and Figure <ref type=\\\"figure\\\" target=\\\"#fig_11\\\">S6B-C</ref>): WM_Astro cells, characterized by high AQP1 expression, are localized primarily to white matter tracts and the peripheral globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3H</ref>); STR_Astro-2 cells, defined by ADGRV1, are enriched in striatum (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3I</ref>), and GP-TH_Astro cells, expressing SLC6A11, are selectively concentrated in the globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3J</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"ADGRV1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Spatial mapping reveal distinct topographic specializations (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3G</ref> and Figure <ref type=\\\"figure\\\" target=\\\"#fig_11\\\">S6B-C</ref>): WM_Astro cells, characterized by high AQP1 expression, are localized primarily to white matter tracts and the peripheral globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3H</ref>); STR_Astro-2 cells, defined by ADGRV1, are enriched in striatum (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3I</ref>), and GP-TH_Astro cells, expressing SLC6A11, are selectively concentrated in the globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3J</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"striatum\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Spatial mapping reveal distinct topographic specializations (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3G</ref> and Figure <ref type=\\\"figure\\\" target=\\\"#fig_11\\\">S6B-C</ref>): WM_Astro cells, characterized by high AQP1 expression, are localized primarily to white matter tracts and the peripheral globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3H</ref>); STR_Astro-2 cells, defined by ADGRV1, are enriched in striatum (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3I</ref>), and GP-TH_Astro cells, expressing SLC6A11, are selectively concentrated in the globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3J</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GP-TH_Astro\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Spatial mapping reveal distinct topographic specializations (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3G</ref> and Figure <ref type=\\\"figure\\\" target=\\\"#fig_11\\\">S6B-C</ref>): WM_Astro cells, characterized by high AQP1 expression, are localized primarily to white matter tracts and the peripheral globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3H</ref>); STR_Astro-2 cells, defined by ADGRV1, are enriched in striatum (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3I</ref>), and GP-TH_Astro cells, expressing SLC6A11, are selectively concentrated in the globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3J</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"SLC6A11\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Spatial mapping reveal distinct topographic specializations (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3G</ref> and Figure <ref type=\\\"figure\\\" target=\\\"#fig_11\\\">S6B-C</ref>): WM_Astro cells, characterized by high AQP1 expression, are localized primarily to white matter tracts and the peripheral globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3H</ref>); STR_Astro-2 cells, defined by ADGRV1, are enriched in striatum (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3I</ref>), and GP-TH_Astro cells, expressing SLC6A11, are selectively concentrated in the globus pallidus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3J</ref>).\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"SLC6A11\",\n      \"label\": \"Gene\",\n      \"sentence\": \"For instance, SLC6A11 (also known as GAT-3) is a GABA transporter that is important for GABA uptake and metabolism in astrocytes <ref type=\\\"bibr\\\" target=\\\"#b49\\\">49,</ref><ref type=\\\"bibr\\\" target=\\\"#b50\\\">50</ref> , consistent with the high abundance of GABAergic neurons in the globus pallidus.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GAT-3\",\n      \"label\": \"Protein\",\n      \"sentence\": \"For instance, SLC6A11 (also known as GAT-3) is a GABA transporter that is important for GABA uptake and metabolism in astrocytes <ref type=\\\"bibr\\\" target=\\\"#b49\\\">49,</ref><ref type=\\\"bibr\\\" target=\\\"#b50\\\">50</ref> , consistent with the high abundance of GABAergic neurons in the globus pallidus.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GABA transporter\",\n      \"label\": \"Protein\",\n      \"sentence\": \"For instance, SLC6A11 (also known as GAT-3) is a GABA transporter that is important for GABA uptake and metabolism in astrocytes <ref type=\\\"bibr\\\" target=\\\"#b49\\\">49,</ref><ref type=\\\"bibr\\\" target=\\\"#b50\\\">50</ref> , consistent with the high abundance of GABAergic neurons in the globus pallidus.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GABA\",\n      \"label\": \"Neurotransmitter\",\n      \"sentence\": \"For instance, SLC6A11 (also known as GAT-3) is a GABA transporter that is important for GABA uptake and metabolism in astrocytes <ref type=\\\"bibr\\\" target=\\\"#b49\\\">49,</ref><ref type=\\\"bibr\\\" target=\\\"#b50\\\">50</ref> , consistent with the high abundance of GABAergic neurons in the globus pallidus.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"astrocytes\",\n      \"label\": \"CellType\",\n      \"sentence\": \"For instance, SLC6A11 (also known as GAT-3) is a GABA transporter that is important for GABA uptake and metabolism in astrocytes <ref type=\\\"bibr\\\" target=\\\"#b49\\\">49,</ref><ref type=\\\"bibr\\\" target=\\\"#b50\\\">50</ref> , consistent with the high abundance of GABAergic neurons in the globus pallidus.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GABAergic neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"For instance, SLC6A11 (also known as GAT-3) is a GABA transporter that is important for GABA uptake and metabolism in astrocytes <ref type=\\\"bibr\\\" target=\\\"#b49\\\">49,</ref><ref type=\\\"bibr\\\" target=\\\"#b50\\\">50</ref> , consistent with the high abundance of GABAergic neurons in the globus pallidus.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"globus pallidus\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"For instance, SLC6A11 (also known as GAT-3) is a GABA transporter that is important for GABA uptake and metabolism in astrocytes <ref type=\\\"bibr\\\" target=\\\"#b49\\\">49,</ref><ref type=\\\"bibr\\\" target=\\\"#b50\\\">50</ref> , consistent with the high abundance of GABAergic neurons in the globus pallidus.\",\n      \"paper_location\": \"Molecular and spatial taxonomy of cell types in the human basal ganglia\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"The high spatial precision of MERFISH+ enables the localization of individual RNA molecules at subcellular resolution (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1J</ref>-K; Supplementary Movie 1).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"RNA molecules\",\n      \"label\": \"NucleicAcid\",\n      \"sentence\": \"The high spatial precision of MERFISH+ enables the localization of individual RNA molecules at subcellular resolution (Figure <ref type=\\\"figure\\\" target=\\\"#fig_14\\\">1J</ref>-K; Supplementary Movie 1).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"nucleus\",\n      \"label\": \"CellularStructure\",\n      \"sentence\": \"This allows each transcript to be categorized as nuclear (within the nucleus), somatic (within the cytoplasm), or putative cellular processes (distributed along processes or extensions) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4A</ref>) (see STAR Methods).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"cytoplasm\",\n      \"label\": \"CellularStructure\",\n      \"sentence\": \"This allows each transcript to be categorized as nuclear (within the nucleus), somatic (within the cytoplasm), or putative cellular processes (distributed along processes or extensions) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4A</ref>) (see STAR Methods).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"putative cellular processes\",\n      \"label\": \"CellularStructure\",\n      \"sentence\": \"This allows each transcript to be categorized as nuclear (within the nucleus), somatic (within the cytoplasm), or putative cellular processes (distributed along processes or extensions) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4A</ref>) (see STAR Methods).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"single neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"While nuclear and somatic transcripts are used to define cell identity, we developed additional analysis to capture somatic morphology of single neurons.\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"human basal ganglia\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"We also used the RNA distributions in putative cellular processes to reveal information about projection topology in human basal ganglia.\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"Cholinergic interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Cholinergic interneurons and globus pallidus neurons including GPi Core, GPi Shell, CN LHX6-GBX1, and GPe MEIS2-SOX6 populations, exhibit the largest median soma volumes (up to ~9,400 µm³), whereas LAMP5+ interneurons and MSNs display markedly smaller volumes (~1,800 µm³ and ~2,500 µm³, respectively) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4B</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"globus pallidus neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Cholinergic interneurons and globus pallidus neurons including GPi Core, GPi Shell, CN LHX6-GBX1, and GPe MEIS2-SOX6 populations, exhibit the largest median soma volumes (up to ~9,400 µm³), whereas LAMP5+ interneurons and MSNs display markedly smaller volumes (~1,800 µm³ and ~2,500 µm³, respectively) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4B</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"GPi Core\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Cholinergic interneurons and globus pallidus neurons including GPi Core, GPi Shell, CN LHX6-GBX1, and GPe MEIS2-SOX6 populations, exhibit the largest median soma volumes (up to ~9,400 µm³), whereas LAMP5+ interneurons and MSNs display markedly smaller volumes (~1,800 µm³ and ~2,500 µm³, respectively) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4B</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"GPi Shell\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Cholinergic interneurons and globus pallidus neurons including GPi Core, GPi Shell, CN LHX6-GBX1, and GPe MEIS2-SOX6 populations, exhibit the largest median soma volumes (up to ~9,400 µm³), whereas LAMP5+ interneurons and MSNs display markedly smaller volumes (~1,800 µm³ and ~2,500 µm³, respectively) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4B</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"CN LHX6-GBX1\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Cholinergic interneurons and globus pallidus neurons including GPi Core, GPi Shell, CN LHX6-GBX1, and GPe MEIS2-SOX6 populations, exhibit the largest median soma volumes (up to ~9,400 µm³), whereas LAMP5+ interneurons and MSNs display markedly smaller volumes (~1,800 µm³ and ~2,500 µm³, respectively) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4B</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"GPe MEIS2-SOX6\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Cholinergic interneurons and globus pallidus neurons including GPi Core, GPi Shell, CN LHX6-GBX1, and GPe MEIS2-SOX6 populations, exhibit the largest median soma volumes (up to ~9,400 µm³), whereas LAMP5+ interneurons and MSNs display markedly smaller volumes (~1,800 µm³ and ~2,500 µm³, respectively) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4B</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"LAMP5+ interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Cholinergic interneurons and globus pallidus neurons including GPi Core, GPi Shell, CN LHX6-GBX1, and GPe MEIS2-SOX6 populations, exhibit the largest median soma volumes (up to ~9,400 µm³), whereas LAMP5+ interneurons and MSNs display markedly smaller volumes (~1,800 µm³ and ~2,500 µm³, respectively) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4B</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"SST-CHODL interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Although MSN subtypes are more consistent in soma volumes, other neuronal classes exhibit more subtype-dependent variability; for example, SST-CHODL interneurons have among the largest median volumes (~6,200 µm³), while SST-ADARB2 interneurons are significantly smaller (~2,300 µm³) (p-value = 6.9 x 10 -15 , Wilcoxon rank-sum test).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"SST-ADARB2 interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Although MSN subtypes are more consistent in soma volumes, other neuronal classes exhibit more subtype-dependent variability; for example, SST-CHODL interneurons have among the largest median volumes (~6,200 µm³), while SST-ADARB2 interneurons are significantly smaller (~2,300 µm³) (p-value = 6.9 x 10 -15 , Wilcoxon rank-sum test).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"Wilcoxon rank-sum test\",\n      \"label\": \"StatisticalTest\",\n      \"sentence\": \"Although MSN subtypes are more consistent in soma volumes, other neuronal classes exhibit more subtype-dependent variability; for example, SST-CHODL interneurons have among the largest median volumes (~6,200 µm³), while SST-ADARB2 interneurons are significantly smaller (~2,300 µm³) (p-value = 6.9 x 10 -15 , Wilcoxon rank-sum test).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"axonal\",\n      \"label\": \"CellularStructure\",\n      \"sentence\": \"We next analyzed RNA distributions in putative cellular processes including axonal and dendritic processes (\\\"neurite transcripts\\\") which provide indirect information about neuronal projections.\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"dendritic processes\",\n      \"label\": \"CellularStructure\",\n      \"sentence\": \"We next analyzed RNA distributions in putative cellular processes including axonal and dendritic processes (\\\"neurite transcripts\\\") which provide indirect information about neuronal projections.\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"neurite transcripts\",\n      \"label\": \"CellularStructure\",\n      \"sentence\": \"We next analyzed RNA distributions in putative cellular processes including axonal and dendritic processes (\\\"neurite transcripts\\\") which provide indirect information about neuronal projections.\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Focusing on MSNs, in the mouse and monkey basal ganglia circuitry, D1 matrix MSNs project directly to the GPi and substantia nigra pars reticulata, whereas D2 matrix MSNs form the indirect pathway via the GPe 3,54 .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"mouse\",\n      \"label\": \"Species\",\n      \"sentence\": \"Focusing on MSNs, in the mouse and monkey basal ganglia circuitry, D1 matrix MSNs project directly to the GPi and substantia nigra pars reticulata, whereas D2 matrix MSNs form the indirect pathway via the GPe 3,54 .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"monkey\",\n      \"label\": \"Species\",\n      \"sentence\": \"Focusing on MSNs, in the mouse and monkey basal ganglia circuitry, D1 matrix MSNs project directly to the GPi and substantia nigra pars reticulata, whereas D2 matrix MSNs form the indirect pathway via the GPe 3,54 .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"basal ganglia circuitry\",\n      \"label\": \"NeuralCircuit\",\n      \"sentence\": \"Focusing on MSNs, in the mouse and monkey basal ganglia circuitry, D1 matrix MSNs project directly to the GPi and substantia nigra pars reticulata, whereas D2 matrix MSNs form the indirect pathway via the GPe 3,54 .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"D1 matrix MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Focusing on MSNs, in the mouse and monkey basal ganglia circuitry, D1 matrix MSNs project directly to the GPi and substantia nigra pars reticulata, whereas D2 matrix MSNs form the indirect pathway via the GPe 3,54 .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"GPi\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Focusing on MSNs, in the mouse and monkey basal ganglia circuitry, D1 matrix MSNs project directly to the GPi and substantia nigra pars reticulata, whereas D2 matrix MSNs form the indirect pathway via the GPe 3,54 .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"substantia nigra pars reticulata\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Focusing on MSNs, in the mouse and monkey basal ganglia circuitry, D1 matrix MSNs project directly to the GPi and substantia nigra pars reticulata, whereas D2 matrix MSNs form the indirect pathway via the GPe 3,54 .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"D2 matrix MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Focusing on MSNs, in the mouse and monkey basal ganglia circuitry, D1 matrix MSNs project directly to the GPi and substantia nigra pars reticulata, whereas D2 matrix MSNs form the indirect pathway via the GPe 3,54 .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"indirect pathway\",\n      \"label\": \"NeuralCircuit\",\n      \"sentence\": \"Focusing on MSNs, in the mouse and monkey basal ganglia circuitry, D1 matrix MSNs project directly to the GPi and substantia nigra pars reticulata, whereas D2 matrix MSNs form the indirect pathway via the GPe 3,54 .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"GPe\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Focusing on MSNs, in the mouse and monkey basal ganglia circuitry, D1 matrix MSNs project directly to the GPi and substantia nigra pars reticulata, whereas D2 matrix MSNs form the indirect pathway via the GPe 3,54 .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"RELN\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Consistent with this organization, we found that D1-enriched transcripts (e.g., RELN, EBF1, SLIT1) exhibit higher extra-somatic signal density within the GPi, while D2-enriched transcripts show preferential accumulation in the GPe (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4C-D</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"EBF1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Consistent with this organization, we found that D1-enriched transcripts (e.g., RELN, EBF1, SLIT1) exhibit higher extra-somatic signal density within the GPi, while D2-enriched transcripts show preferential accumulation in the GPe (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4C-D</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"SLIT1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Consistent with this organization, we found that D1-enriched transcripts (e.g., RELN, EBF1, SLIT1) exhibit higher extra-somatic signal density within the GPi, while D2-enriched transcripts show preferential accumulation in the GPe (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4C-D</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"GPi\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Consistent with this organization, we found that D1-enriched transcripts (e.g., RELN, EBF1, SLIT1) exhibit higher extra-somatic signal density within the GPi, while D2-enriched transcripts show preferential accumulation in the GPe (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4C-D</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"GPe\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Consistent with this organization, we found that D1-enriched transcripts (e.g., RELN, EBF1, SLIT1) exhibit higher extra-somatic signal density within the GPi, while D2-enriched transcripts show preferential accumulation in the GPe (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4C-D</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"Quantitative correlation analysis across the 673 genes imaged with MERFISH+ confirms this relationship: D1specific genes display significantly higher spatial correlation scores in the GPi than in the GPe, whereas the inverse pattern is observed for D2-specific genes (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4E</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"D1specific genes\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Quantitative correlation analysis across the 673 genes imaged with MERFISH+ confirms this relationship: D1specific genes display significantly higher spatial correlation scores in the GPi than in the GPe, whereas the inverse pattern is observed for D2-specific genes (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4E</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"GPi\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Quantitative correlation analysis across the 673 genes imaged with MERFISH+ confirms this relationship: D1specific genes display significantly higher spatial correlation scores in the GPi than in the GPe, whereas the inverse pattern is observed for D2-specific genes (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4E</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"GPe\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Quantitative correlation analysis across the 673 genes imaged with MERFISH+ confirms this relationship: D1specific genes display significantly higher spatial correlation scores in the GPi than in the GPe, whereas the inverse pattern is observed for D2-specific genes (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4E</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"D2-specific genes\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Quantitative correlation analysis across the 673 genes imaged with MERFISH+ confirms this relationship: D1specific genes display significantly higher spatial correlation scores in the GPi than in the GPe, whereas the inverse pattern is observed for D2-specific genes (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4E</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"SEMA3E\",\n      \"label\": \"Gene\",\n      \"sentence\": \"In addition, we observed that GPi and GPe regions have high concentration of extrasomatic SEMA3E and SEMA5A, ligands part of the semaphorin family of axonal guidance molecules <ref type=\\\"bibr\\\" target=\\\"#b55\\\">55,</ref><ref type=\\\"bibr\\\" target=\\\"#b56\\\">56</ref> .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"SEMA5A\",\n      \"label\": \"Gene\",\n      \"sentence\": \"In addition, we observed that GPi and GPe regions have high concentration of extrasomatic SEMA3E and SEMA5A, ligands part of the semaphorin family of axonal guidance molecules <ref type=\\\"bibr\\\" target=\\\"#b55\\\">55,</ref><ref type=\\\"bibr\\\" target=\\\"#b56\\\">56</ref> .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"semaphorin family\",\n      \"label\": \"ProteinFamily\",\n      \"sentence\": \"In addition, we observed that GPi and GPe regions have high concentration of extrasomatic SEMA3E and SEMA5A, ligands part of the semaphorin family of axonal guidance molecules <ref type=\\\"bibr\\\" target=\\\"#b55\\\">55,</ref><ref type=\\\"bibr\\\" target=\\\"#b56\\\">56</ref> .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"axonal guidance molecules\",\n      \"label\": \"Protein\",\n      \"sentence\": \"In addition, we observed that GPi and GPe regions have high concentration of extrasomatic SEMA3E and SEMA5A, ligands part of the semaphorin family of axonal guidance molecules <ref type=\\\"bibr\\\" target=\\\"#b55\\\">55,</ref><ref type=\\\"bibr\\\" target=\\\"#b56\\\">56</ref> .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"SEMA3E\",\n      \"label\": \"Gene\",\n      \"sentence\": \"We found SEMA3E is enriched in GPe (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4F</ref>), while SEMA5A is prominently found in GPi (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4H</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"GPe\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"We found SEMA3E is enriched in GPe (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4F</ref>), while SEMA5A is prominently found in GPi (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4H</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"SEMA5A\",\n      \"label\": \"Gene\",\n      \"sentence\": \"We found SEMA3E is enriched in GPe (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4F</ref>), while SEMA5A is prominently found in GPi (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4H</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"GPi\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"We found SEMA3E is enriched in GPe (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4F</ref>), while SEMA5A is prominently found in GPi (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4H</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"snRNA-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"Their expression is also higher in the nuclei as confirmed using the snRNA-seq data from the Allen Institute Basal Ganglia Consensus Cell Type Atlas (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4G</ref>, <ref type=\\\"figure\\\">I</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"Allen Institute Basal Ganglia Consensus Cell Type Atlas\",\n      \"label\": \"DataResource\",\n      \"sentence\": \"Their expression is also higher in the nuclei as confirmed using the snRNA-seq data from the Allen Institute Basal Ganglia Consensus Cell Type Atlas (Figure <ref type=\\\"figure\\\" target=\\\"#fig_5\\\">4G</ref>, <ref type=\\\"figure\\\">I</ref>).\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"This might reflect that specific semaphorin gene family members mediate long range connections in neurons such as MSNs targeting specific segments of the globus pallidus <ref type=\\\"bibr\\\" target=\\\"#b55\\\">55,</ref><ref type=\\\"bibr\\\" target=\\\"#b57\\\">57</ref> .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"This might reflect that specific semaphorin gene family members mediate long range connections in neurons such as MSNs targeting specific segments of the globus pallidus <ref type=\\\"bibr\\\" target=\\\"#b55\\\">55,</ref><ref type=\\\"bibr\\\" target=\\\"#b57\\\">57</ref> .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"globus pallidus\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"This might reflect that specific semaphorin gene family members mediate long range connections in neurons such as MSNs targeting specific segments of the globus pallidus <ref type=\\\"bibr\\\" target=\\\"#b55\\\">55,</ref><ref type=\\\"bibr\\\" target=\\\"#b57\\\">57</ref> .\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"Together, these findings demonstrate that MERFISH+ not only resolves molecular and spatial identities of cells but also captures morphological and projection-inferred features encoded in the subcellular transcriptome and provides an unprecedented window into the structural and functional organization of human basal ganglia circuits.\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"human basal ganglia circuits\",\n      \"label\": \"NeuralCircuit\",\n      \"sentence\": \"Together, these findings demonstrate that MERFISH+ not only resolves molecular and spatial identities of cells but also captures morphological and projection-inferred features encoded in the subcellular transcriptome and provides an unprecedented window into the structural and functional organization of human basal ganglia circuits.\",\n      \"paper_location\": \"Subcellular transcript localization for somatic size estimation and projection inference\"\n    },\n    {\n      \"entity\": \"cellular community analysis\",\n      \"label\": \"Method\",\n      \"sentence\": \"To resolve mesoscale organization beyond individual cell types, we performed cellular community analysis <ref type=\\\"bibr\\\" target=\\\"#b13\\\">14,</ref><ref type=\\\"bibr\\\" target=\\\"#b58\\\">58</ref> , which groups cells based on the composition of their spatial neighborhoods.\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"Allen Human Brain Atlas\",\n      \"label\": \"DataResource\",\n      \"sentence\": \"This approach yields 14 discrete spatial modules that correspond to anatomical regions defined in the Allen Human Brain Atlas <ref type=\\\"bibr\\\" target=\\\"#b59\\\">59</ref> (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5B</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GPe SOX8-CTXND1\",\n      \"label\": \"CellType\",\n      \"sentence\": \"For example, GPe SOX8-CTXND1 and GPe MEIS2-SOX6 neurons are enriched in the external globus pallidus (GPe), whereas GPi Core and GPi Shell neurons are confined to the GPi.\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GPe MEIS2-SOX6 neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"For example, GPe SOX8-CTXND1 and GPe MEIS2-SOX6 neurons are enriched in the external globus pallidus (GPe), whereas GPi Core and GPi Shell neurons are confined to the GPi.\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"external globus pallidus\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"For example, GPe SOX8-CTXND1 and GPe MEIS2-SOX6 neurons are enriched in the external globus pallidus (GPe), whereas GPi Core and GPi Shell neurons are confined to the GPi.\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GPe\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"For example, GPe SOX8-CTXND1 and GPe MEIS2-SOX6 neurons are enriched in the external globus pallidus (GPe), whereas GPi Core and GPi Shell neurons are confined to the GPi.\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GPi Core\",\n      \"label\": \"CellType\",\n      \"sentence\": \"For example, GPe SOX8-CTXND1 and GPe MEIS2-SOX6 neurons are enriched in the external globus pallidus (GPe), whereas GPi Core and GPi Shell neurons are confined to the GPi.\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GPi Shell neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"For example, GPe SOX8-CTXND1 and GPe MEIS2-SOX6 neurons are enriched in the external globus pallidus (GPe), whereas GPi Core and GPi Shell neurons are confined to the GPi.\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GPi\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"For example, GPe SOX8-CTXND1 and GPe MEIS2-SOX6 neurons are enriched in the external globus pallidus (GPe), whereas GPi Core and GPi Shell neurons are confined to the GPi.\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRv D1\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Similarly, STRv D1 and STRv D2 MSNs are enriched in the nucleus accumbens shell and core, while STRd D1 and STRd D2 MSNs populate the striatum (caudate nucleus and putamen) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5C</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRv D2 MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Similarly, STRv D1 and STRv D2 MSNs are enriched in the nucleus accumbens shell and core, while STRd D1 and STRd D2 MSNs populate the striatum (caudate nucleus and putamen) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5C</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"nucleus accumbens shell\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Similarly, STRv D1 and STRv D2 MSNs are enriched in the nucleus accumbens shell and core, while STRd D1 and STRd D2 MSNs populate the striatum (caudate nucleus and putamen) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5C</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRd D1\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Similarly, STRv D1 and STRv D2 MSNs are enriched in the nucleus accumbens shell and core, while STRd D1 and STRd D2 MSNs populate the striatum (caudate nucleus and putamen) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5C</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRd D2 MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Similarly, STRv D1 and STRv D2 MSNs are enriched in the nucleus accumbens shell and core, while STRd D1 and STRd D2 MSNs populate the striatum (caudate nucleus and putamen) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5C</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"caudate nucleus\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Similarly, STRv D1 and STRv D2 MSNs are enriched in the nucleus accumbens shell and core, while STRd D1 and STRd D2 MSNs populate the striatum (caudate nucleus and putamen) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5C</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"putamen\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Similarly, STRv D1 and STRv D2 MSNs are enriched in the nucleus accumbens shell and core, while STRd D1 and STRd D2 MSNs populate the striatum (caudate nucleus and putamen) (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5C</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STR fast-spiking PTHLH-PVALB GABAergic interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"For instance, STR fast-spiking PTHLH-PVALB GABAergic interneurons are preferentially enriched in the caudate nucleus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5D</ref>, in agreement with previous snRNA-seq analysis <ref type=\\\"bibr\\\" target=\\\"#b10\\\">11</ref> .\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"caudate nucleus\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"For instance, STR fast-spiking PTHLH-PVALB GABAergic interneurons are preferentially enriched in the caudate nucleus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5D</ref>, in agreement with previous snRNA-seq analysis <ref type=\\\"bibr\\\" target=\\\"#b10\\\">11</ref> .\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"snRNA-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"For instance, STR fast-spiking PTHLH-PVALB GABAergic interneurons are preferentially enriched in the caudate nucleus (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5D</ref>, in agreement with previous snRNA-seq analysis <ref type=\\\"bibr\\\" target=\\\"#b10\\\">11</ref> .\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"STRd ChAT+ GABAergic interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"STRd ChAT+ GABAergic interneurons are more abundant in the putamen (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5E</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"putamen\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"STRd ChAT+ GABAergic interneurons are more abundant in the putamen (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5E</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"FM GATA3-TCF7L2 neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"In contrast, FM GATA3-TCF7L2 neurons, a transcriptionally distinct population with unclear lineage affiliation are enriched in the GPi nucleus of the basal ganglia; ; this represents a previously unknown neuronal subtype within this region (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5F</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"GPi nucleus\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"In contrast, FM GATA3-TCF7L2 neurons, a transcriptionally distinct population with unclear lineage affiliation are enriched in the GPi nucleus of the basal ganglia; ; this represents a previously unknown neuronal subtype within this region (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5F</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"D1-and D2-striosomal MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"In contrast, both D1-and D2-striosomal MSNs are detected in the NAc shell but are largely absent from the NAc core (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5G-H</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"NAc shell\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"In contrast, both D1-and D2-striosomal MSNs are detected in the NAc shell but are largely absent from the NAc core (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5G-H</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"NAc core\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"In contrast, both D1-and D2-striosomal MSNs are detected in the NAc shell but are largely absent from the NAc core (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5G-H</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"TH\",\n      \"label\": \"Gene\",\n      \"sentence\": \"This differential organization is corroborated by the expression of established molecular markers: TH associated with striosome compartments <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3</ref> , is enriched in the NAc shell, whereas KIRREL3, a matrixassociated marker <ref type=\\\"bibr\\\" target=\\\"#b60\\\">60</ref> , is predominantly expressed in the NAc core (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5I</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"striosome compartments\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"This differential organization is corroborated by the expression of established molecular markers: TH associated with striosome compartments <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3</ref> , is enriched in the NAc shell, whereas KIRREL3, a matrixassociated marker <ref type=\\\"bibr\\\" target=\\\"#b60\\\">60</ref> , is predominantly expressed in the NAc core (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5I</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"NAc shell\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"This differential organization is corroborated by the expression of established molecular markers: TH associated with striosome compartments <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3</ref> , is enriched in the NAc shell, whereas KIRREL3, a matrixassociated marker <ref type=\\\"bibr\\\" target=\\\"#b60\\\">60</ref> , is predominantly expressed in the NAc core (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5I</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"KIRREL3\",\n      \"label\": \"Gene\",\n      \"sentence\": \"This differential organization is corroborated by the expression of established molecular markers: TH associated with striosome compartments <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3</ref> , is enriched in the NAc shell, whereas KIRREL3, a matrixassociated marker <ref type=\\\"bibr\\\" target=\\\"#b60\\\">60</ref> , is predominantly expressed in the NAc core (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5I</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"NAc core\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"This differential organization is corroborated by the expression of established molecular markers: TH associated with striosome compartments <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3</ref> , is enriched in the NAc shell, whereas KIRREL3, a matrixassociated marker <ref type=\\\"bibr\\\" target=\\\"#b60\\\">60</ref> , is predominantly expressed in the NAc core (Figure <ref type=\\\"figure\\\" target=\\\"#fig_6\\\">5I</ref>).\",\n      \"paper_location\": \"Mesoscale cellular community architecture of the human basal ganglia\"\n    },\n    {\n      \"entity\": \"striatum\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Beyond its mesoscale structures, the striatum exhibits micro-domain architecture composed of striosome and matrix compartments <ref type=\\\"bibr\\\" target=\\\"#b61\\\">61</ref> .\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"striosome\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"Beyond its mesoscale structures, the striatum exhibits micro-domain architecture composed of striosome and matrix compartments <ref type=\\\"bibr\\\" target=\\\"#b61\\\">61</ref> .\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"matrix compartments\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"Beyond its mesoscale structures, the striatum exhibits micro-domain architecture composed of striosome and matrix compartments <ref type=\\\"bibr\\\" target=\\\"#b61\\\">61</ref> .\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"MSN\",\n      \"label\": \"CellType\",\n      \"sentence\": \"To delineate these domains, we further applied cellular community analysis restricted to MSN subtype composition-the principal determinant of striosome-matrix identity (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6A-B</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"STR TAC3-PLPP4 interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Notably, STR TAC3-PLPP4 interneurons are exclusively localized to the matrix compartment (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6C</ref>, 6D top).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"matrix compartment\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"Notably, STR TAC3-PLPP4 interneurons are exclusively localized to the matrix compartment (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6C</ref>, 6D top).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"Tac2+/TH+ interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"These interneurons are primate-specific <ref type=\\\"bibr\\\" target=\\\"#b62\\\">62,</ref><ref type=\\\"bibr\\\" target=\\\"#b63\\\">63</ref> and likely represent the orthologous lineage of rare Tac2+/TH+ interneurons that is described in rodents <ref type=\\\"bibr\\\" target=\\\"#b16\\\">17,</ref><ref type=\\\"bibr\\\" target=\\\"#b64\\\">64</ref> .\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"rodents\",\n      \"label\": \"Species\",\n      \"sentence\": \"These interneurons are primate-specific <ref type=\\\"bibr\\\" target=\\\"#b62\\\">62,</ref><ref type=\\\"bibr\\\" target=\\\"#b63\\\">63</ref> and likely represent the orthologous lineage of rare Tac2+/TH+ interneurons that is described in rodents <ref type=\\\"bibr\\\" target=\\\"#b16\\\">17,</ref><ref type=\\\"bibr\\\" target=\\\"#b64\\\">64</ref> .\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"STR TAC3-PLPP4 neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"In our human dataset, STR TAC3-PLPP4 neurons are among the most abundant population of striatal interneurons (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6E</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"striatal interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"In our human dataset, STR TAC3-PLPP4 neurons are among the most abundant population of striatal interneurons (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6E</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"TAC3-PLPP4 interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"This marked expansion of TAC3-PLPP4 interneurons in primates suggests an evolutionary adaptation that supports the integration of the greater density of prefrontal and associative cortical inputs, potentially enabling enhanced sensorimotor and cognitive control <ref type=\\\"bibr\\\" target=\\\"#b10\\\">11,</ref><ref type=\\\"bibr\\\" target=\\\"#b65\\\">65</ref> .\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"primates\",\n      \"label\": \"Species\",\n      \"sentence\": \"This marked expansion of TAC3-PLPP4 interneurons in primates suggests an evolutionary adaptation that supports the integration of the greater density of prefrontal and associative cortical inputs, potentially enabling enhanced sensorimotor and cognitive control <ref type=\\\"bibr\\\" target=\\\"#b10\\\">11,</ref><ref type=\\\"bibr\\\" target=\\\"#b65\\\">65</ref> .\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"prefrontal\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"This marked expansion of TAC3-PLPP4 interneurons in primates suggests an evolutionary adaptation that supports the integration of the greater density of prefrontal and associative cortical inputs, potentially enabling enhanced sensorimotor and cognitive control <ref type=\\\"bibr\\\" target=\\\"#b10\\\">11,</ref><ref type=\\\"bibr\\\" target=\\\"#b65\\\">65</ref> .\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"VIP+\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Conversely, VIP+ and LAMP5+ interneurons display moderate enrichment within striosomal territories (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6C</ref>, 6D bottom), although these populations are not exclusive to that compartment.\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"LAMP5+ interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Conversely, VIP+ and LAMP5+ interneurons display moderate enrichment within striosomal territories (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6C</ref>, 6D bottom), although these populations are not exclusive to that compartment.\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"striosomal territories\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"Conversely, VIP+ and LAMP5+ interneurons display moderate enrichment within striosomal territories (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6C</ref>, 6D bottom), although these populations are not exclusive to that compartment.\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"astrocyte\",\n      \"label\": \"CellType\",\n      \"sentence\": \"This analysis reveals astrocyte and Oligo OPALIN oligodendrocyte enrichment near striosome borders that form boundary-aligned microdomains (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6G</ref>, 6H bottom).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"Oligo OPALIN oligodendrocyte\",\n      \"label\": \"CellType\",\n      \"sentence\": \"This analysis reveals astrocyte and Oligo OPALIN oligodendrocyte enrichment near striosome borders that form boundary-aligned microdomains (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6G</ref>, 6H bottom).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"striosome borders\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"This analysis reveals astrocyte and Oligo OPALIN oligodendrocyte enrichment near striosome borders that form boundary-aligned microdomains (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6G</ref>, 6H bottom).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"Oligo OPALIN subtype\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The Oligo OPALIN subtype, previously identified along striatopallidal fiber tracts (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3</ref>), appears to delineate the perimeters of striosomal territories and highlights glial contributions to compartmental architecture.\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"striatopallidal fiber tracts\",\n      \"label\": \"NeuralCircuit\",\n      \"sentence\": \"The Oligo OPALIN subtype, previously identified along striatopallidal fiber tracts (Figure <ref type=\\\"figure\\\" target=\\\"#fig_4\\\">3</ref>), appears to delineate the perimeters of striosomal territories and highlights glial contributions to compartmental architecture.\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"striosomes\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"To assess whether gene expression varies systematically with spatial proximity to striosomes, we examined transcript gradients along the striosome-to-matrix axis.\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"striosome-to-matrix axis\",\n      \"label\": \"AnatomicalAxis\",\n      \"sentence\": \"To assess whether gene expression varies systematically with spatial proximity to striosomes, we examined transcript gradients along the striosome-to-matrix axis.\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"astrocyte-enriched genes\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Several astrocyte-enriched genes exhibit distance-dependent expression patterns (Figure <ref type=\\\"figure\\\" target=\\\"#fig_8\\\">S7A</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"Allen Institute's whole-mouse-brain MERFISH dataset\",\n      \"label\": \"DataResource\",\n      \"sentence\": \"We next investigated cross-species conservation of this organization using the Allen Institute's whole-mouse-brain MERFISH dataset <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10</ref> .\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"MERFISH\",\n      \"label\": \"Method\",\n      \"sentence\": \"We next investigated cross-species conservation of this organization using the Allen Institute's whole-mouse-brain MERFISH dataset <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10</ref> .\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"Orthologous mouse MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Orthologous mouse MSNs were identified by correlating gene-expression signatures with human subtypes (Figure <ref type=\\\"figure\\\">S8A</ref>, <ref type=\\\"figure\\\">B</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"Mouse supertypes 0268 STR D1 GABA_4\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Mouse supertypes 0268 STR D1 GABA_4 and 0267 STR D1 GABA_3 correspond to human STRd D1 Striosome and STRd D1 Matrix MSNs, respectively (Figure <ref type=\\\"figure\\\">S8C-D</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"0267 STR D1 GABA_3\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Mouse supertypes 0268 STR D1 GABA_4 and 0267 STR D1 GABA_3 correspond to human STRd D1 Striosome and STRd D1 Matrix MSNs, respectively (Figure <ref type=\\\"figure\\\">S8C-D</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"human STRd D1 Striosome\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Mouse supertypes 0268 STR D1 GABA_4 and 0267 STR D1 GABA_3 correspond to human STRd D1 Striosome and STRd D1 Matrix MSNs, respectively (Figure <ref type=\\\"figure\\\">S8C-D</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"STRd D1 Matrix MSNs\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Mouse supertypes 0268 STR D1 GABA_4 and 0267 STR D1 GABA_3 correspond to human STRd D1 Striosome and STRd D1 Matrix MSNs, respectively (Figure <ref type=\\\"figure\\\">S8C-D</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"0277 STR D2 Gaba_4\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Similarly, finer subclustering of the mouse 0277 STR D2 Gaba_4 supertype, designated as A* and B*, map to human STRd D2 Striosome MSN and STRd D2 Matrix MSN, respectively (Figure <ref type=\\\"figure\\\">S8E-F</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"human STRd D2 Striosome MSN\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Similarly, finer subclustering of the mouse 0277 STR D2 Gaba_4 supertype, designated as A* and B*, map to human STRd D2 Striosome MSN and STRd D2 Matrix MSN, respectively (Figure <ref type=\\\"figure\\\">S8E-F</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"STRd D2 Matrix MSN\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Similarly, finer subclustering of the mouse 0277 STR D2 Gaba_4 supertype, designated as A* and B*, map to human STRd D2 Striosome MSN and STRd D2 Matrix MSN, respectively (Figure <ref type=\\\"figure\\\">S8E-F</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"mouse striatum\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Using these orthologous MSN pairs, we delineated striosome-matrix boundaries in the mouse striatum and performed analogous proximity analyses in which we computed the distance of cells from border of nearest striosome (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6I</ref>) and quantified the distribution of cells as a function of distance.\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"Astro-TE NN_3 astrocytes\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The results are highly concordant across species: specific glial populations, Astro-TE NN_3 astrocytes and MOL NN_4 oligodendrocytes, are enriched at striosome borders in mice (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6J-K</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"MOL NN_4 oligodendrocytes\",\n      \"label\": \"CellType\",\n      \"sentence\": \"The results are highly concordant across species: specific glial populations, Astro-TE NN_3 astrocytes and MOL NN_4 oligodendrocytes, are enriched at striosome borders in mice (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6J-K</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"mice\",\n      \"label\": \"Species\",\n      \"sentence\": \"The results are highly concordant across species: specific glial populations, Astro-TE NN_3 astrocytes and MOL NN_4 oligodendrocytes, are enriched at striosome borders in mice (Figure <ref type=\\\"figure\\\" target=\\\"#fig_7\\\">6J-K</ref>).\",\n      \"paper_location\": \"Striosome-matrix domains and cross-species orthologous organization\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"To directly examine these patterns, we applied unsupervised clustering to spatially weighted whole-transcriptome profiles derived from Stereo-seq data, independent of prior cell-type annotation (see Methods).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"This spatial analysis recapitulates the mesoscale modules identified by MERFISH+ cellular community analysis and aligns closely with established anatomical territories of the human basal ganglia (Figure <ref type=\\\"figure\\\" target=\\\"#fig_8\\\">7A</ref>, Figure <ref type=\\\"figure\\\">S9A</ref>).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"human basal ganglia\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"This spatial analysis recapitulates the mesoscale modules identified by MERFISH+ cellular community analysis and aligns closely with established anatomical territories of the human basal ganglia (Figure <ref type=\\\"figure\\\" target=\\\"#fig_8\\\">7A</ref>, Figure <ref type=\\\"figure\\\">S9A</ref>).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"putamen\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"For example, the putamen exhibits strong enrichment of ZNF385B (a zinc-finger transcription factor) and ROBO2 (an axonal guidance molecule), whereas the caudate nucleus shows elevated expression of TAC1(a neuropeptide precursor) and EGR1(an immediate early transcription factor).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"ZNF385B\",\n      \"label\": \"Gene\",\n      \"sentence\": \"For example, the putamen exhibits strong enrichment of ZNF385B (a zinc-finger transcription factor) and ROBO2 (an axonal guidance molecule), whereas the caudate nucleus shows elevated expression of TAC1(a neuropeptide precursor) and EGR1(an immediate early transcription factor).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"ROBO2\",\n      \"label\": \"Gene\",\n      \"sentence\": \"For example, the putamen exhibits strong enrichment of ZNF385B (a zinc-finger transcription factor) and ROBO2 (an axonal guidance molecule), whereas the caudate nucleus shows elevated expression of TAC1(a neuropeptide precursor) and EGR1(an immediate early transcription factor).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"caudate nucleus\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"For example, the putamen exhibits strong enrichment of ZNF385B (a zinc-finger transcription factor) and ROBO2 (an axonal guidance molecule), whereas the caudate nucleus shows elevated expression of TAC1(a neuropeptide precursor) and EGR1(an immediate early transcription factor).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"TAC1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"For example, the putamen exhibits strong enrichment of ZNF385B (a zinc-finger transcription factor) and ROBO2 (an axonal guidance molecule), whereas the caudate nucleus shows elevated expression of TAC1(a neuropeptide precursor) and EGR1(an immediate early transcription factor).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"EGR1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"For example, the putamen exhibits strong enrichment of ZNF385B (a zinc-finger transcription factor) and ROBO2 (an axonal guidance molecule), whereas the caudate nucleus shows elevated expression of TAC1(a neuropeptide precursor) and EGR1(an immediate early transcription factor).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"nucleus accumbens shell\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"The nucleus accumbens shell is marked by vesicle associated molecules SCG2 and CALY, while the NAc core is enriched for stress-response genes HSPA6 and DNAJB1 (Figure <ref type=\\\"figure\\\" target=\\\"#fig_8\\\">7C</ref>).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"SCG2\",\n      \"label\": \"Gene\",\n      \"sentence\": \"The nucleus accumbens shell is marked by vesicle associated molecules SCG2 and CALY, while the NAc core is enriched for stress-response genes HSPA6 and DNAJB1 (Figure <ref type=\\\"figure\\\" target=\\\"#fig_8\\\">7C</ref>).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"CALY\",\n      \"label\": \"Gene\",\n      \"sentence\": \"The nucleus accumbens shell is marked by vesicle associated molecules SCG2 and CALY, while the NAc core is enriched for stress-response genes HSPA6 and DNAJB1 (Figure <ref type=\\\"figure\\\" target=\\\"#fig_8\\\">7C</ref>).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"NAc core\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"The nucleus accumbens shell is marked by vesicle associated molecules SCG2 and CALY, while the NAc core is enriched for stress-response genes HSPA6 and DNAJB1 (Figure <ref type=\\\"figure\\\" target=\\\"#fig_8\\\">7C</ref>).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"HSPA6\",\n      \"label\": \"Gene\",\n      \"sentence\": \"The nucleus accumbens shell is marked by vesicle associated molecules SCG2 and CALY, while the NAc core is enriched for stress-response genes HSPA6 and DNAJB1 (Figure <ref type=\\\"figure\\\" target=\\\"#fig_8\\\">7C</ref>).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"DNAJB1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"The nucleus accumbens shell is marked by vesicle associated molecules SCG2 and CALY, while the NAc core is enriched for stress-response genes HSPA6 and DNAJB1 (Figure <ref type=\\\"figure\\\" target=\\\"#fig_8\\\">7C</ref>).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"gene ontology\",\n      \"label\": \"Method\",\n      \"sentence\": \"We further examined the transcriptional difference between the regions by performing gene ontology (GO) enrichment analysis on the differentially expressed genes.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"GO\",\n      \"label\": \"Method\",\n      \"sentence\": \"We further examined the transcriptional difference between the regions by performing gene ontology (GO) enrichment analysis on the differentially expressed genes.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"putamen\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Pathways such as synaptic signaling and regulation of cell communication are mainly associated with putamen and NAc shell, while processes such as regulation of secretion by cell and regulation of response to stress prominently involve the caudate with some contribution from the NAc shell and NAc core respectively.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"NAc shell\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Pathways such as synaptic signaling and regulation of cell communication are mainly associated with putamen and NAc shell, while processes such as regulation of secretion by cell and regulation of response to stress prominently involve the caudate with some contribution from the NAc shell and NAc core respectively.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"caudate\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Pathways such as synaptic signaling and regulation of cell communication are mainly associated with putamen and NAc shell, while processes such as regulation of secretion by cell and regulation of response to stress prominently involve the caudate with some contribution from the NAc shell and NAc core respectively.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"NAc core\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Pathways such as synaptic signaling and regulation of cell communication are mainly associated with putamen and NAc shell, while processes such as regulation of secretion by cell and regulation of response to stress prominently involve the caudate with some contribution from the NAc shell and NAc core respectively.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"In our Stereo-seq data, the genes displaying regional specificity constitute only ~5%; majority of the remaining genes are expressed throughout the striatum.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"striatum\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"In our Stereo-seq data, the genes displaying regional specificity constitute only ~5%; majority of the remaining genes are expressed throughout the striatum.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"mice\",\n      \"label\": \"Species\",\n      \"sentence\": \"This led us to ask whether gene expression gradients in the human striatum are similar to those known in mice and non-human primates <ref type=\\\"bibr\\\" target=\\\"#b3\\\">4,</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">5,</ref><ref type=\\\"bibr\\\" target=\\\"#b66\\\">66</ref> .\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"non-human primates\",\n      \"label\": \"Species\",\n      \"sentence\": \"This led us to ask whether gene expression gradients in the human striatum are similar to those known in mice and non-human primates <ref type=\\\"bibr\\\" target=\\\"#b3\\\">4,</ref><ref type=\\\"bibr\\\" target=\\\"#b4\\\">5,</ref><ref type=\\\"bibr\\\" target=\\\"#b66\\\">66</ref> .\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"dorsolateral-ventromedial\",\n      \"label\": \"AnatomicalAxis\",\n      \"sentence\": \"To investigate topographical continuity of gene expression across species we focused on dorsolateral-ventromedial (DL-VM) expression gradients, an established organizational axis in the mouse striatum <ref type=\\\"bibr\\\" target=\\\"#b13\\\">14,</ref><ref type=\\\"bibr\\\" target=\\\"#b66\\\">66</ref> .\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"DL-VM\",\n      \"label\": \"AnatomicalAxis\",\n      \"sentence\": \"To investigate topographical continuity of gene expression across species we focused on dorsolateral-ventromedial (DL-VM) expression gradients, an established organizational axis in the mouse striatum <ref type=\\\"bibr\\\" target=\\\"#b13\\\">14,</ref><ref type=\\\"bibr\\\" target=\\\"#b66\\\">66</ref> .\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"mouse striatum\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"To investigate topographical continuity of gene expression across species we focused on dorsolateral-ventromedial (DL-VM) expression gradients, an established organizational axis in the mouse striatum <ref type=\\\"bibr\\\" target=\\\"#b13\\\">14,</ref><ref type=\\\"bibr\\\" target=\\\"#b66\\\">66</ref> .\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"PDE10A\",\n      \"label\": \"Gene\",\n      \"sentence\": \"For example, PDE10A, NRGN, and SLIT1 are dorsolaterally enriched, whereas NNAT, CRYM, and RAP1GAP exhibit ventromedial bias.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"NRGN\",\n      \"label\": \"Gene\",\n      \"sentence\": \"For example, PDE10A, NRGN, and SLIT1 are dorsolaterally enriched, whereas NNAT, CRYM, and RAP1GAP exhibit ventromedial bias.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"SLIT1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"For example, PDE10A, NRGN, and SLIT1 are dorsolaterally enriched, whereas NNAT, CRYM, and RAP1GAP exhibit ventromedial bias.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"NNAT\",\n      \"label\": \"Gene\",\n      \"sentence\": \"For example, PDE10A, NRGN, and SLIT1 are dorsolaterally enriched, whereas NNAT, CRYM, and RAP1GAP exhibit ventromedial bias.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"CRYM\",\n      \"label\": \"Gene\",\n      \"sentence\": \"For example, PDE10A, NRGN, and SLIT1 are dorsolaterally enriched, whereas NNAT, CRYM, and RAP1GAP exhibit ventromedial bias.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"RAP1GAP\",\n      \"label\": \"Gene\",\n      \"sentence\": \"For example, PDE10A, NRGN, and SLIT1 are dorsolaterally enriched, whereas NNAT, CRYM, and RAP1GAP exhibit ventromedial bias.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"Allen Institute Basal Ganglia Consensus Cell Type Atlas\",\n      \"label\": \"DataResource\",\n      \"sentence\": \"To identify cellular contributors to these spatial gradients, we mapped gradientassociated genes onto cell types defined in the Allen Institute Basal Ganglia Consensus Cell Type Atlas.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"lateral ganglionic eminence\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Most gradient-associated transcripts are enriched in lateral ganglionic eminence (CN/LGE)-derived GABAergic neurons, the principal MSN population.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"CN/LGE\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Most gradient-associated transcripts are enriched in lateral ganglionic eminence (CN/LGE)-derived GABAergic neurons, the principal MSN population.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"GABAergic neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Most gradient-associated transcripts are enriched in lateral ganglionic eminence (CN/LGE)-derived GABAergic neurons, the principal MSN population.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"MSN population\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Most gradient-associated transcripts are enriched in lateral ganglionic eminence (CN/LGE)-derived GABAergic neurons, the principal MSN population.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"oligodendrocyte lineage\",\n      \"label\": \"CellType\",\n      \"sentence\": \"A subset of dorsolateral-enriched genes, however, are expressed within oligodendrocyte lineage (OPC-Oligo) cells, that reflects the dense myelinated fiber tracts that occupy dorsal striatal territories (Figure <ref type=\\\"figure\\\" target=\\\"#fig_8\\\">7I</ref>).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"OPC-Oligo\",\n      \"label\": \"CellType\",\n      \"sentence\": \"A subset of dorsolateral-enriched genes, however, are expressed within oligodendrocyte lineage (OPC-Oligo) cells, that reflects the dense myelinated fiber tracts that occupy dorsal striatal territories (Figure <ref type=\\\"figure\\\" target=\\\"#fig_8\\\">7I</ref>).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"myelinated fiber tracts\",\n      \"label\": \"NeuralCircuit\",\n      \"sentence\": \"A subset of dorsolateral-enriched genes, however, are expressed within oligodendrocyte lineage (OPC-Oligo) cells, that reflects the dense myelinated fiber tracts that occupy dorsal striatal territories (Figure <ref type=\\\"figure\\\" target=\\\"#fig_8\\\">7I</ref>).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"LGE neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"In contrast, in mouse we do not observe this; LGE neurons dominate in both directions (Figure <ref type=\\\"figure\\\">S9E</ref>).\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"These analyses demonstrate that Stereo-seq captures transcriptome-wide spatial gradients that mirror functional and anatomical axes of the basal ganglia.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"basal ganglia\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"These analyses demonstrate that Stereo-seq captures transcriptome-wide spatial gradients that mirror functional and anatomical axes of the basal ganglia.\",\n      \"paper_location\": \"Stereo-seq mapping\"\n    },\n    {\n      \"entity\": \"Allen Human Brain Atlas\",\n      \"label\": \"DataResource\",\n      \"sentence\": \"Our cellular community analysis based on the proximity of the 60 identified cell types reveals 14 spatial modules in the human basal ganglia sections, providing an unbiased, refined anatomical delineation of the imaged regions that aligns closely with the Allen Human Brain Atlas parcellation <ref type=\\\"bibr\\\" target=\\\"#b59\\\">59</ref> .\",\n      \"paper_location\": \"DISCUSSION: Mesoscale cellular communities and striosome-matrix architecture\"\n    },\n    {\n      \"entity\": \"Mammalian Basal Ganglia Consensus Cell Type Atlas\",\n      \"label\": \"DataResource\",\n      \"sentence\": \"The differential enrichment of cell-types across spatial modules recapitulates expected spatial segregations based on the Mammalian Basal Ganglia Consensus Cell Type Atlas reference (Allen Institute; RRID:SCR_024672)\",\n      \"paper_location\": \"DISCUSSION: Mesoscale cellular communities and striosome-matrix architecture\"\n    },\n    {\n      \"entity\": \"ChAT+\",\n      \"label\": \"CellType\",\n      \"sentence\": \"in NAc; STRd D1/D2 in the caudate/putamen) including regional interneuron biases: STR FS PTHLH-PVALB enriched in the caudate, ChAT+ enriched in the putamen, and FM GATA3-TCF7L2 concentrated in GPi.\",\n      \"paper_location\": \"DISCUSSION: Mesoscale cellular communities and striosome-matrix architecture\"\n    },\n    {\n      \"entity\": \"TAC3-PLPP4 interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Specifically, TAC3-PLPP4 interneurons, which are expanded in primates <ref type=\\\"bibr\\\" target=\\\"#b64\\\">64</ref> , are matrix-restricted, whereas VIP+ and LAMP5+ interneurons are striosome-enriched but not exclusive.\",\n      \"paper_location\": \"DISCUSSION: Mesoscale cellular communities and striosome-matrix architecture\"\n    },\n    {\n      \"entity\": \"primates\",\n      \"label\": \"Species\",\n      \"sentence\": \"Specifically, TAC3-PLPP4 interneurons, which are expanded in primates <ref type=\\\"bibr\\\" target=\\\"#b64\\\">64</ref> , are matrix-restricted, whereas VIP+ and LAMP5+ interneurons are striosome-enriched but not exclusive.\",\n      \"paper_location\": \"DISCUSSION: Mesoscale cellular communities and striosome-matrix architecture\"\n    },\n    {\n      \"entity\": \"VIP+\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Specifically, TAC3-PLPP4 interneurons, which are expanded in primates <ref type=\\\"bibr\\\" target=\\\"#b64\\\">64</ref> , are matrix-restricted, whereas VIP+ and LAMP5+ interneurons are striosome-enriched but not exclusive.\",\n      \"paper_location\": \"DISCUSSION: Mesoscale cellular communities and striosome-matrix architecture\"\n    },\n    {\n      \"entity\": \"LAMP5+ interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Specifically, TAC3-PLPP4 interneurons, which are expanded in primates <ref type=\\\"bibr\\\" target=\\\"#b64\\\">64</ref> , are matrix-restricted, whereas VIP+ and LAMP5+ interneurons are striosome-enriched but not exclusive.\",\n      \"paper_location\": \"DISCUSSION: Mesoscale cellular communities and striosome-matrix architecture\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"MERFISH+ subcellular transcript localizations enable characterization of cell morphology and projection-inferred connectivity.\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"immunofluorescence staining\",\n      \"label\": \"Method\",\n      \"sentence\": \"Although this strategy differs from traditional methods that offer more detailed cell morphology, including immunofluorescence staining and Cajal-Golgi impregnation methods <ref type=\\\"bibr\\\" target=\\\"#b69\\\">69,</ref><ref type=\\\"bibr\\\">70</ref> , the key advantage of MERFISH+ based quantification is its ability to distinguish and measure dozens of neuronal subclasses at tissue-wide scale.\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"Cajal-Golgi impregnation methods\",\n      \"label\": \"Method\",\n      \"sentence\": \"Although this strategy differs from traditional methods that offer more detailed cell morphology, including immunofluorescence staining and Cajal-Golgi impregnation methods <ref type=\\\"bibr\\\" target=\\\"#b69\\\">69,</ref><ref type=\\\"bibr\\\">70</ref> , the key advantage of MERFISH+ based quantification is its ability to distinguish and measure dozens of neuronal subclasses at tissue-wide scale.\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"Although this strategy differs from traditional methods that offer more detailed cell morphology, including immunofluorescence staining and Cajal-Golgi impregnation methods <ref type=\\\"bibr\\\" target=\\\"#b69\\\">69,</ref><ref type=\\\"bibr\\\">70</ref> , the key advantage of MERFISH+ based quantification is its ability to distinguish and measure dozens of neuronal subclasses at tissue-wide scale.\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"D1-enriched genes\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Specifically, D1-enriched genes (e.g., RELN, EBF1, SLIT1) are found to preferentially accumulate in GPi projections, whereas D2-enriched genes are biased to GPe projections.\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"RELN\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Specifically, D1-enriched genes (e.g., RELN, EBF1, SLIT1) are found to preferentially accumulate in GPi projections, whereas D2-enriched genes are biased to GPe projections.\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"EBF1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Specifically, D1-enriched genes (e.g., RELN, EBF1, SLIT1) are found to preferentially accumulate in GPi projections, whereas D2-enriched genes are biased to GPe projections.\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"SLIT1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Specifically, D1-enriched genes (e.g., RELN, EBF1, SLIT1) are found to preferentially accumulate in GPi projections, whereas D2-enriched genes are biased to GPe projections.\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"GPi projections\",\n      \"label\": \"NeuralCircuit\",\n      \"sentence\": \"Specifically, D1-enriched genes (e.g., RELN, EBF1, SLIT1) are found to preferentially accumulate in GPi projections, whereas D2-enriched genes are biased to GPe projections.\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"D2-enriched genes\",\n      \"label\": \"Gene\",\n      \"sentence\": \"Specifically, D1-enriched genes (e.g., RELN, EBF1, SLIT1) are found to preferentially accumulate in GPi projections, whereas D2-enriched genes are biased to GPe projections.\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"GPe projections\",\n      \"label\": \"NeuralCircuit\",\n      \"sentence\": \"Specifically, D1-enriched genes (e.g., RELN, EBF1, SLIT1) are found to preferentially accumulate in GPi projections, whereas D2-enriched genes are biased to GPe projections.\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"direct (D1→GPi) versus indirect (D2→GPe) pathway topology\",\n      \"label\": \"NeuralCircuit\",\n      \"sentence\": \"These correlation patterns recapitulate the direct (D1→GPi) versus indirect (D2→GPe) pathway topology in rodent and non-human primate models <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6,</ref><ref type=\\\"bibr\\\" target=\\\"#b68\\\">68</ref> .\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"rodent\",\n      \"label\": \"Species\",\n      \"sentence\": \"These correlation patterns recapitulate the direct (D1→GPi) versus indirect (D2→GPe) pathway topology in rodent and non-human primate models <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6,</ref><ref type=\\\"bibr\\\" target=\\\"#b68\\\">68</ref> .\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"non-human primate\",\n      \"label\": \"Species\",\n      \"sentence\": \"These correlation patterns recapitulate the direct (D1→GPi) versus indirect (D2→GPe) pathway topology in rodent and non-human primate models <ref type=\\\"bibr\\\" target=\\\"#b2\\\">3,</ref><ref type=\\\"bibr\\\" target=\\\"#b5\\\">6,</ref><ref type=\\\"bibr\\\" target=\\\"#b68\\\">68</ref> .\",\n      \"paper_location\": \"DISCUSSION: Subcellular transcript localization informs morphology and projection architecture\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"Unsupervised clustering of the whole-transcriptome spatial distribution (based on Stereo-seq) recapitulates MERFISH+ cell-type community modules and classical anatomy and provides novel gene enrichment within each anatomical subregions.\",\n      \"paper_location\": \"DISCUSSION: Genome-wide transcriptional modules and a conserved DL-VM gradient\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"Unsupervised clustering of the whole-transcriptome spatial distribution (based on Stereo-seq) recapitulates MERFISH+ cell-type community modules and classical anatomy and provides novel gene enrichment within each anatomical subregions.\",\n      \"paper_location\": \"DISCUSSION: Genome-wide transcriptional modules and a conserved DL-VM gradient\"\n    },\n    {\n      \"entity\": \"dorsolateral-ventromedial gradient\",\n      \"label\": \"AnatomicalAxis\",\n      \"sentence\": \"We find that a dorsolateral-ventromedial gradient emerges as a dominant axis of transcriptional variation at the multi-centimeter-scale; this axis is conserved in the mouse at the multi-millimeter scale <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">14</ref> revealing dorsolateral-biased genes such as PDE10A, NRGN and SLIT1 and ventromedial-biased genes such as NNAT, CRYM, RAP1GAP.\",\n      \"paper_location\": \"DISCUSSION: Genome-wide transcriptional modules and a conserved DL-VM gradient\"\n    },\n    {\n      \"entity\": \"PDE10A\",\n      \"label\": \"Gene\",\n      \"sentence\": \"We find that a dorsolateral-ventromedial gradient emerges as a dominant axis of transcriptional variation at the multi-centimeter-scale; this axis is conserved in the mouse at the multi-millimeter scale <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">14</ref> revealing dorsolateral-biased genes such as PDE10A, NRGN and SLIT1 and ventromedial-biased genes such as NNAT, CRYM, RAP1GAP.\",\n      \"paper_location\": \"DISCUSSION: Genome-wide transcriptional modules and a conserved DL-VM gradient\"\n    },\n    {\n      \"entity\": \"NRGN\",\n      \"label\": \"Gene\",\n      \"sentence\": \"We find that a dorsolateral-ventromedial gradient emerges as a dominant axis of transcriptional variation at the multi-centimeter-scale; this axis is conserved in the mouse at the multi-millimeter scale <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">14</ref> revealing dorsolateral-biased genes such as PDE10A, NRGN and SLIT1 and ventromedial-biased genes such as NNAT, CRYM, RAP1GAP.\",\n      \"paper_location\": \"DISCUSSION: Genome-wide transcriptional modules and a conserved DL-VM gradient\"\n    },\n    {\n      \"entity\": \"SLIT1\",\n      \"label\": \"Gene\",\n      \"sentence\": \"We find that a dorsolateral-ventromedial gradient emerges as a dominant axis of transcriptional variation at the multi-centimeter-scale; this axis is conserved in the mouse at the multi-millimeter scale <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">14</ref> revealing dorsolateral-biased genes such as PDE10A, NRGN and SLIT1 and ventromedial-biased genes such as NNAT, CRYM, RAP1GAP.\",\n      \"paper_location\": \"DISCUSSION: Genome-wide transcriptional modules and a conserved DL-VM gradient\"\n    },\n    {\n      \"entity\": \"NNAT\",\n      \"label\": \"Gene\",\n      \"sentence\": \"We find that a dorsolateral-ventromedial gradient emerges as a dominant axis of transcriptional variation at the multi-centimeter-scale; this axis is conserved in the mouse at the multi-millimeter scale <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">14</ref> revealing dorsolateral-biased genes such as PDE10A, NRGN and SLIT1 and ventromedial-biased genes such as NNAT, CRYM, RAP1GAP.\",\n      \"paper_location\": \"DISCUSSION: Genome-wide transcriptional modules and a conserved DL-VM gradient\"\n    },\n    {\n      \"entity\": \"CRYM\",\n      \"label\": \"Gene\",\n      \"sentence\": \"We find that a dorsolateral-ventromedial gradient emerges as a dominant axis of transcriptional variation at the multi-centimeter-scale; this axis is conserved in the mouse at the multi-millimeter scale <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">14</ref> revealing dorsolateral-biased genes such as PDE10A, NRGN and SLIT1 and ventromedial-biased genes such as NNAT, CRYM, RAP1GAP.\",\n      \"paper_location\": \"DISCUSSION: Genome-wide transcriptional modules and a conserved DL-VM gradient\"\n    },\n    {\n      \"entity\": \"RAP1GAP\",\n      \"label\": \"Gene\",\n      \"sentence\": \"We find that a dorsolateral-ventromedial gradient emerges as a dominant axis of transcriptional variation at the multi-centimeter-scale; this axis is conserved in the mouse at the multi-millimeter scale <ref type=\\\"bibr\\\" target=\\\"#b9\\\">10,</ref><ref type=\\\"bibr\\\" target=\\\"#b13\\\">14</ref> revealing dorsolateral-biased genes such as PDE10A, NRGN and SLIT1 and ventromedial-biased genes such as NNAT, CRYM, RAP1GAP.\",\n      \"paper_location\": \"DISCUSSION: Genome-wide transcriptional modules and a conserved DL-VM gradient\"\n    },\n    {\n      \"entity\": \"CN LGE GABA neurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Most of the genes exhibiting the gradient are expressed in CN LGE GABA neurons, with a dorsal contribution from OPC-Oligo lineages, consistent with differential myelination in the dorsal fiber bundles <ref type=\\\"bibr\\\">73,</ref><ref type=\\\"bibr\\\">74</ref> .\",\n      \"paper_location\": \"DISCUSSION: Genome-wide transcriptional modules and a conserved DL-VM gradient\"\n    },\n    {\n      \"entity\": \"OPC-Oligo lineages\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Most of the genes exhibiting the gradient are expressed in CN LGE GABA neurons, with a dorsal contribution from OPC-Oligo lineages, consistent with differential myelination in the dorsal fiber bundles <ref type=\\\"bibr\\\">73,</ref><ref type=\\\"bibr\\\">74</ref> .\",\n      \"paper_location\": \"DISCUSSION: Genome-wide transcriptional modules and a conserved DL-VM gradient\"\n    },\n    {\n      \"entity\": \"dorsal fiber bundles\",\n      \"label\": \"NeuralCircuit\",\n      \"sentence\": \"Most of the genes exhibiting the gradient are expressed in CN LGE GABA neurons, with a dorsal contribution from OPC-Oligo lineages, consistent with differential myelination in the dorsal fiber bundles <ref type=\\\"bibr\\\">73,</ref><ref type=\\\"bibr\\\">74</ref> .\",\n      \"paper_location\": \"DISCUSSION: Genome-wide transcriptional modules and a conserved DL-VM gradient\"\n    },\n    {\n      \"entity\": \"neurotypical human basal ganglia\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Functionally, the underlying cell types, cellular communities, subcellular transcript geography, and conserved spatial gradients suggest an intricate multi-layered organizational architecture for neurotypical human basal ganglia <ref type=\\\"bibr\\\">75,</ref><ref type=\\\"bibr\\\">76</ref> .\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"subcellular RNA localization\",\n      \"label\": \"Method\",\n      \"sentence\": \"The finding that subcellular RNA localization captures projection-inferred signatures suggests a promising avenue for inferring human circuit topology directly from spatial transcriptomic data -an approach that circumvents the limitations of traditional neural tracing in human tissue.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"human circuit topology\",\n      \"label\": \"NeuralCircuit\",\n      \"sentence\": \"The finding that subcellular RNA localization captures projection-inferred signatures suggests a promising avenue for inferring human circuit topology directly from spatial transcriptomic data -an approach that circumvents the limitations of traditional neural tracing in human tissue.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"neural tracing\",\n      \"label\": \"Method\",\n      \"sentence\": \"The finding that subcellular RNA localization captures projection-inferred signatures suggests a promising avenue for inferring human circuit topology directly from spatial transcriptomic data -an approach that circumvents the limitations of traditional neural tracing in human tissue.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"astrocyte\",\n      \"label\": \"CellType\",\n      \"sentence\": \"In addition, astrocyte and oligodendrocyte niches aligned with striosome boundaries point to specialized metabolic and myelination interfaces at these compartmental perimeters.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"oligodendrocyte\",\n      \"label\": \"CellType\",\n      \"sentence\": \"In addition, astrocyte and oligodendrocyte niches aligned with striosome boundaries point to specialized metabolic and myelination interfaces at these compartmental perimeters.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"striosome boundaries\",\n      \"label\": \"AnatomicalCompartment\",\n      \"sentence\": \"In addition, astrocyte and oligodendrocyte niches aligned with striosome boundaries point to specialized metabolic and myelination interfaces at these compartmental perimeters.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"TAC3-PLPP4 interneurons\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Finally, the primate-expanded, matrix-restricted TAC3-PLPP4 interneurons may reflect evolutionary adaptations to the enlarged prefrontal cortex, potentially supporting the increased computational demands of human-specific action selection, valuation, and cognitive flexibility <ref type=\\\"bibr\\\" target=\\\"#b10\\\">11</ref> .\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"prefrontal cortex\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Finally, the primate-expanded, matrix-restricted TAC3-PLPP4 interneurons may reflect evolutionary adaptations to the enlarged prefrontal cortex, potentially supporting the increased computational demands of human-specific action selection, valuation, and cognitive flexibility <ref type=\\\"bibr\\\" target=\\\"#b10\\\">11</ref> .\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"GPi core\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Clinically, our dataset highlights specific neuronal and glial subtypes including GPi core versus shell populations and boundary-enriched glial communities, as potential substrates of selective vulnerability in movement and neuropsychiatric disorders.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"shell populations\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Clinically, our dataset highlights specific neuronal and glial subtypes including GPi core versus shell populations and boundary-enriched glial communities, as potential substrates of selective vulnerability in movement and neuropsychiatric disorders.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"glial communities\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Clinically, our dataset highlights specific neuronal and glial subtypes including GPi core versus shell populations and boundary-enriched glial communities, as potential substrates of selective vulnerability in movement and neuropsychiatric disorders.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"neuropsychiatric disorders\",\n      \"label\": \"Disease\",\n      \"sentence\": \"Clinically, our dataset highlights specific neuronal and glial subtypes including GPi core versus shell populations and boundary-enriched glial communities, as potential substrates of selective vulnerability in movement and neuropsychiatric disorders.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"D1\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Establishing the neurotypical abundance and spatial distribution of D1, D2, and D1/D2 hybrid MSN subtypes in the caudate, putamen, and nucleus accumbens provides a crucial baseline for detecting disease-associated shifts in cell number, composition, and transcriptional state.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"D2\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Establishing the neurotypical abundance and spatial distribution of D1, D2, and D1/D2 hybrid MSN subtypes in the caudate, putamen, and nucleus accumbens provides a crucial baseline for detecting disease-associated shifts in cell number, composition, and transcriptional state.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"D1/D2 hybrid MSN subtypes\",\n      \"label\": \"CellType\",\n      \"sentence\": \"Establishing the neurotypical abundance and spatial distribution of D1, D2, and D1/D2 hybrid MSN subtypes in the caudate, putamen, and nucleus accumbens provides a crucial baseline for detecting disease-associated shifts in cell number, composition, and transcriptional state.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"caudate\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Establishing the neurotypical abundance and spatial distribution of D1, D2, and D1/D2 hybrid MSN subtypes in the caudate, putamen, and nucleus accumbens provides a crucial baseline for detecting disease-associated shifts in cell number, composition, and transcriptional state.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"putamen\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Establishing the neurotypical abundance and spatial distribution of D1, D2, and D1/D2 hybrid MSN subtypes in the caudate, putamen, and nucleus accumbens provides a crucial baseline for detecting disease-associated shifts in cell number, composition, and transcriptional state.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"nucleus accumbens\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"Establishing the neurotypical abundance and spatial distribution of D1, D2, and D1/D2 hybrid MSN subtypes in the caudate, putamen, and nucleus accumbens provides a crucial baseline for detecting disease-associated shifts in cell number, composition, and transcriptional state.\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"Huntington's disease\",\n      \"label\": \"Disease\",\n      \"sentence\": \"Such comparisons will be particularly valuable for elucidating celltype-selective vulnerability in basal ganglia disorders, including Huntington's disease and Parkinson's disease <ref type=\\\"bibr\\\">77,</ref><ref type=\\\"bibr\\\">78</ref> .\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"Parkinson's disease\",\n      \"label\": \"Disease\",\n      \"sentence\": \"Such comparisons will be particularly valuable for elucidating celltype-selective vulnerability in basal ganglia disorders, including Huntington's disease and Parkinson's disease <ref type=\\\"bibr\\\">77,</ref><ref type=\\\"bibr\\\">78</ref> .\",\n      \"paper_location\": \"Functional implications\"\n    },\n    {\n      \"entity\": \"postmortem human tissue\",\n      \"label\": \"TissueType\",\n      \"sentence\": \"This study is limited by the number and demographic diversity of available donors, as well as inherent constraints associated with postmortem human tissue.\",\n      \"paper_location\": \"Limitations of the study\"\n    },\n    {\n      \"entity\": \"immunohistochemistry\",\n      \"label\": \"Method\",\n      \"sentence\": \"Although our multimodal spatial framework provides a comprehensive baseline, additional orthogonal validations such as immunohistochemistry, in situ sequencing, and in vivo electrophysiological or connectivity assays would further strengthen cell-type and spatial community assignments.\",\n      \"paper_location\": \"Limitations of the study\"\n    },\n    {\n      \"entity\": \"in situ sequencing\",\n      \"label\": \"Method\",\n      \"sentence\": \"Although our multimodal spatial framework provides a comprehensive baseline, additional orthogonal validations such as immunohistochemistry, in situ sequencing, and in vivo electrophysiological or connectivity assays would further strengthen cell-type and spatial community assignments.\",\n      \"paper_location\": \"Limitations of the study\"\n    },\n    {\n      \"entity\": \"in vivo electrophysiological\",\n      \"label\": \"Method\",\n      \"sentence\": \"Although our multimodal spatial framework provides a comprehensive baseline, additional orthogonal validations such as immunohistochemistry, in situ sequencing, and in vivo electrophysiological or connectivity assays would further strengthen cell-type and spatial community assignments.\",\n      \"paper_location\": \"Limitations of the study\"\n    },\n    {\n      \"entity\": \"connectivity assays\",\n      \"label\": \"Method\",\n      \"sentence\": \"Although our multimodal spatial framework provides a comprehensive baseline, additional orthogonal validations such as immunohistochemistry, in situ sequencing, and in vivo electrophysiological or connectivity assays would further strengthen cell-type and spatial community assignments.\",\n      \"paper_location\": \"Limitations of the study\"\n    },\n    {\n      \"entity\": \"neurological and psychiatric disorders\",\n      \"label\": \"Disease\",\n      \"sentence\": \"Future work should extend this approach to a broader range of ages and demographic backgrounds and apply this methodology to comparative analyses of neurological and psychiatric disorders in human tissue.\",\n      \"paper_location\": \"Limitations of the study\"\n    },\n    {\n      \"entity\": \"human tissue\",\n      \"label\": \"TissueType\",\n      \"sentence\": \"Future work should extend this approach to a broader range of ages and demographic backgrounds and apply this methodology to comparative analyses of neurological and psychiatric disorders in human tissue.\",\n      \"paper_location\": \"Limitations of the study\"\n    },\n    {\n      \"entity\": \"human basal ganglia\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"In summary, by coupling genome-wide spatial profiling with subcellular transcript localization and community-level modeling, we deliver a multiscale, molecular atlas of human basal ganglia organization from single transcripts to cellular communities and anatomical modules that shows features unique to humans and those that are conserved across species.\",\n      \"paper_location\": \"Limitations of the study\"\n    },\n    {\n      \"entity\": \"neuropsychiatric and neurodegenerative disorders\",\n      \"label\": \"Disease\",\n      \"sentence\": \"This atlas establishes a foundational framework for studying human basal ganglia architecture and its disruption in neuropsychiatric and neurodegenerative disorders, enabling cross-species and disease-state comparisons at single-cell and community scales.\",\n      \"paper_location\": \"Limitations of the study\"\n    },\n    {\n      \"entity\": \"Healthy adult basal ganglia brain tissue\",\n      \"label\": \"TissueType\",\n      \"sentence\": \"Biological samples Healthy adult basal ganglia brain tissue University of California Irvine BICAN Brain Bank Case#UCI-2724 Case#UCI-3924 Case#UCI-1311 Case#UCI-5129\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"Isopentane\",\n      \"label\": \"Chemical\",\n      \"sentence\": \"Chemicals, peptides, and recombinant proteins Isopentane Sigma-Aldrich Cat#M33631-4L Methanol Sigma-Aldrich Cat#1060071000 Ethanol (100%) Decon Labs Cat# V1016 Formamide Ambion (Thermo Fisher Scientific) Cat# AM9342 TE Buffer, pH 8.0 Thermo Fisher Scientific Cat# T11493 20X SSC Thermo Fisher Scientific Cat# J60839.K3 Triton X-100 Sigma-Aldrich Cat# 92046-34-9 Trolox Sigma-Aldrich Cat# 238813 Poly-D-lysine Sigma-Aldrich Cat# A-003-M Gel Slick Lonza Cat# 50640 PFA (4%) Thermo Fisher Scientific Cat# 28906 DAPI Thermo Fisher Scientific Cat# D1306 10% SDS Corning Cat# 46-040-Cl TEMED Sigma-Aldrich Cat# T7024 Catalase Sigma-Aldrich Cat# C30-500MG Glucose Oxidase Sigma-Aldrich Cat# G2133-250KU\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"Methanol\",\n      \"label\": \"Chemical\",\n      \"sentence\": \"Chemicals, peptides, and recombinant proteins Isopentane Sigma-Aldrich Cat#M33631-4L Methanol Sigma-Aldrich Cat#1060071000 Ethanol (100%) Decon Labs Cat# V1016 Formamide Ambion (Thermo Fisher Scientific) Cat# AM9342 TE Buffer, pH 8.0 Thermo Fisher Scientific Cat# T11493 20X SSC Thermo Fisher Scientific Cat# J60839.K3 Triton X-100 Sigma-Aldrich Cat# 92046-34-9 Trolox Sigma-Aldrich Cat# 238813 Poly-D-lysine Sigma-Aldrich Cat# A-003-M Gel Slick Lonza Cat# 50640 PFA (4%) Thermo Fisher Scientific Cat# 28906 DAPI Thermo Fisher Scientific Cat# D1306 10% SDS Corning Cat# 46-040-Cl TEMED Sigma-Aldrich Cat# T7024 Catalase Sigma-Aldrich Cat# C30-500MG Glucose Oxidase Sigma-Aldrich Cat# G2133-250KU\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"Ethanol\",\n      \"label\": \"Chemical\",\n      \"sentence\": \"Chemicals, peptides, and recombinant proteins Isopentane Sigma-Aldrich Cat#M33631-4L Methanol Sigma-Aldrich Cat#1060071000 Ethanol (100%) Decon Labs Cat# V1016 Formamide Ambion (Thermo Fisher Scientific) Cat# AM9342 TE Buffer, pH 8.0 Thermo Fisher Scientific Cat# T11493 20X SSC Thermo Fisher Scientific Cat# J60839.K3 Triton X-100 Sigma-Aldrich Cat# 92046-34-9 Trolox Sigma-Aldrich Cat# 238813 Poly-D-lysine Sigma-Aldrich Cat# A-003-M Gel Slick Lonza Cat# 50640 PFA (4%) Thermo Fisher Scientific Cat# 28906 DAPI Thermo Fisher Scientific Cat# D1306 10% SDS Corning Cat# 46-040-Cl TEMED Sigma-Aldrich Cat# T7024 Catalase Sigma-Aldrich Cat# C30-500MG Glucose Oxidase Sigma-Aldrich Cat# G2133-250KU\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"Formamide\",\n      \"label\": \"Chemical\",\n      \"sentence\": \"Chemicals, peptides, and recombinant proteins Isopentane Sigma-Aldrich Cat#M33631-4L Methanol Sigma-Aldrich Cat#1060071000 Ethanol (100%) Decon Labs Cat# V1016 Formamide Ambion (Thermo Fisher Scientific) Cat# AM9342 TE Buffer, pH 8.0 Thermo Fisher Scientific Cat# T11493 20X SSC Thermo Fisher Scientific Cat# J60839.K3 Triton X-100 Sigma-Aldrich Cat# 92046-34-9 Trolox Sigma-Aldrich Cat# 238813 Poly-D-lysine Sigma-Aldrich Cat# A-003-M Gel Slick Lonza Cat# 50640 PFA (4%) Thermo Fisher Scientific Cat# 28906 DAPI Thermo Fisher Scientific Cat# D1306 10% SDS Corning Cat# 46-040-Cl TEMED Sigma-Aldrich Cat# T7024 Catalase Sigma-Aldrich Cat# C30-500MG Glucose Oxidase Sigma-Aldrich Cat# G2133-250KU\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"Triton X-100\",\n      \"label\": \"Chemical\",\n      \"sentence\": \"Chemicals, peptides, and recombinant proteins Isopentane Sigma-Aldrich Cat#M33631-4L Methanol Sigma-Aldrich Cat#1060071000 Ethanol (100%) Decon Labs Cat# V1016 Formamide Ambion (Thermo Fisher Scientific) Cat# AM9342 TE Buffer, pH 8.0 Thermo Fisher Scientific Cat# T11493 20X SSC Thermo Fisher Scientific Cat# J60839.K3 Triton X-100 Sigma-Aldrich Cat# 92046-34-9 Trolox Sigma-Aldrich Cat# 238813 Poly-D-lysine Sigma-Aldrich Cat# A-003-M Gel Slick Lonza Cat# 50640 PFA (4%) Thermo Fisher Scientific Cat# 28906 DAPI Thermo Fisher Scientific Cat# D1306 10% SDS Corning Cat# 46-040-Cl TEMED Sigma-Aldrich Cat# T7024 Catalase Sigma-Aldrich Cat# C30-500MG Glucose Oxidase Sigma-Aldrich Cat# G2133-250KU\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"DAPI\",\n      \"label\": \"Chemical\",\n      \"sentence\": \"Chemicals, peptides, and recombinant proteins Isopentane Sigma-Aldrich Cat#M33631-4L Methanol Sigma-Aldrich Cat#1060071000 Ethanol (100%) Decon Labs Cat# V1016 Formamide Ambion (Thermo Fisher Scientific) Cat# AM9342 TE Buffer, pH 8.0 Thermo Fisher Scientific Cat# T11493 20X SSC Thermo Fisher Scientific Cat# J60839.K3 Triton X-100 Sigma-Aldrich Cat# 92046-34-9 Trolox Sigma-Aldrich Cat# 238813 Poly-D-lysine Sigma-Aldrich Cat# A-003-M Gel Slick Lonza Cat# 50640 PFA (4%) Thermo Fisher Scientific Cat# 28906 DAPI Thermo Fisher Scientific Cat# D1306 10% SDS Corning Cat# 46-040-Cl TEMED Sigma-Aldrich Cat# T7024 Catalase Sigma-Aldrich Cat# C30-500MG Glucose Oxidase Sigma-Aldrich Cat# G2133-250KU\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"Maxima H Minus Reverse Transcriptase\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Maxima H Minus Reverse Transcriptase Thermo Fisher Scientific Cat# EP0753 Phusion High-Fidelity DNA Polymerase New England Biolabs Cat# M0536L HiScribe T7 Quick High Yield RNA Synthesis Kit New England Biolabs Cat# E2050 RNase Inhibitor New England Biolabs Cat# M0314L dNTP Mix (10 mM) Thermo Fisher Scientific Cat# R0192\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"Phusion High-Fidelity DNA Polymerase\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Maxima H Minus Reverse Transcriptase Thermo Fisher Scientific Cat# EP0753 Phusion High-Fidelity DNA Polymerase New England Biolabs Cat# M0536L HiScribe T7 Quick High Yield RNA Synthesis Kit New England Biolabs Cat# E2050 RNase Inhibitor New England Biolabs Cat# M0314L dNTP Mix (10 mM) Thermo Fisher Scientific Cat# R0192\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"RNase Inhibitor\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Maxima H Minus Reverse Transcriptase Thermo Fisher Scientific Cat# EP0753 Phusion High-Fidelity DNA Polymerase New England Biolabs Cat# M0536L HiScribe T7 Quick High Yield RNA Synthesis Kit New England Biolabs Cat# E2050 RNase Inhibitor New England Biolabs Cat# M0314L dNTP Mix (10 mM) Thermo Fisher Scientific Cat# R0192\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"Stereo-seq Permeabilization Set\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"Stereo-seq Permeabilization Set for Chip-on-a-slide V1.1 Complete Genomics Cat#211SP11118-CG Stereo-seq Transcriptomics Set (2cm*3cm) V1.2 STOmics/Complete Genomics Cat#111ST13231-CG Stereo-seq 16 Barcoded Library Preparation Kit V1.0 Complete Genomics Cat#111KL160 PE100 flow cells reagent Complete Genomics Cat#100008555\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"Stereo-seq Transcriptomics Set\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"Stereo-seq Permeabilization Set for Chip-on-a-slide V1.1 Complete Genomics Cat#211SP11118-CG Stereo-seq Transcriptomics Set (2cm*3cm) V1.2 STOmics/Complete Genomics Cat#111ST13231-CG Stereo-seq 16 Barcoded Library Preparation Kit V1.0 Complete Genomics Cat#111KL160 PE100 flow cells reagent Complete Genomics Cat#100008555\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"Stereo-seq 16 Barcoded Library Preparation Kit\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"Stereo-seq Permeabilization Set for Chip-on-a-slide V1.1 Complete Genomics Cat#211SP11118-CG Stereo-seq Transcriptomics Set (2cm*3cm) V1.2 STOmics/Complete Genomics Cat#111ST13231-CG Stereo-seq 16 Barcoded Library Preparation Kit V1.0 Complete Genomics Cat#111KL160 PE100 flow cells reagent Complete Genomics Cat#100008555\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"Encoding probes\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"Oligonucleotides Encoding probes targeting 673 genes Twist Bioscience This paper; Table S1 Readout probes Integrated DNA Technologies (IDT) This paper; Supplemental Data Primers for MERFISH encoding probe synthesis Integrated DNA Technologies (IDT) This paper; Supplemental Data RT primer for cDNA synthesis in RT&TAMP Integrated DNA Technologies (IDT) This paper; Supplemental Data\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"Readout probes\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"Oligonucleotides Encoding probes targeting 673 genes Twist Bioscience This paper; Table S1 Readout probes Integrated DNA Technologies (IDT) This paper; Supplemental Data Primers for MERFISH encoding probe synthesis Integrated DNA Technologies (IDT) This paper; Supplemental Data RT primer for cDNA synthesis in RT&TAMP Integrated DNA Technologies (IDT) This paper; Supplemental Data\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"MERFISH\",\n      \"label\": \"Method\",\n      \"sentence\": \"Oligonucleotides Encoding probes targeting 673 genes Twist Bioscience This paper; Table S1 Readout probes Integrated DNA Technologies (IDT) This paper; Supplemental Data Primers for MERFISH encoding probe synthesis Integrated DNA Technologies (IDT) This paper; Supplemental Data RT primer for cDNA synthesis in RT&TAMP Integrated DNA Technologies (IDT) This paper; Supplemental Data\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"Storm-control\",\n      \"label\": \"Software\",\n      \"sentence\": \"Software & Instrument Storm-control Harvard University Zhuang Lab (<ref type=\\\"url\\\" target=\\\"https://github.com/ZhuangLa\\\">https://github.com/ZhuangLa</ref> b/storm-control) RRID: SCR_027698\",\n      \"paper_location\": \"Key Resources Table\"\n    },\n    {\n      \"entity\": \"basal ganglia\",\n      \"label\": \"BrainRegion\",\n      \"sentence\": \"To facilitate multimodal investigation of the same donor sample, each pre-dissected postmortem human coronal brain slab, which is 5-10 mm thick and containing the basal ganglia, was used in this study.\",\n      \"paper_location\": \"Preparation of multi-centimeter tissue sections for Stereo-seq and MERFISH+\"\n    },\n    {\n      \"entity\": \"OCT embedding medium\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"To prepare adjacent brain sections for experiments, a tissue block measuring up to 4 cm × 5 cm and encompassing basal ganglia components was further dissected out of each coronal brain slab, quickly embedded in OCT embedding medium, and frozen in dry ice prechilled isopentane (SigmaAldrich, M33631).\",\n      \"paper_location\": \"Preparation of multi-centimeter tissue sections for Stereo-seq and MERFISH+\"\n    },\n    {\n      \"entity\": \"isopentane\",\n      \"label\": \"Chemical\",\n      \"sentence\": \"To prepare adjacent brain sections for experiments, a tissue block measuring up to 4 cm × 5 cm and encompassing basal ganglia components was further dissected out of each coronal brain slab, quickly embedded in OCT embedding medium, and frozen in dry ice prechilled isopentane (SigmaAldrich, M33631).\",\n      \"paper_location\": \"Preparation of multi-centimeter tissue sections for Stereo-seq and MERFISH+\"\n    },\n    {\n      \"entity\": \"Leica CM1850 cryostat\",\n      \"label\": \"Instrument\",\n      \"sentence\": \"Cryosections of 10 µm and 16 µm thickness were then obtained using a Leica CM1850 cryostat at -18 to -20 °C and mounted onto Stereo-seq chips and specially coated 50 mm × 64 mm rectangular glass slides for MERFISH+, respectively.\",\n      \"paper_location\": \"Preparation of multi-centimeter tissue sections for Stereo-seq and MERFISH+\"\n    },\n    {\n      \"entity\": \"Stereo-seq chips\",\n      \"label\": \"Instrument\",\n      \"sentence\": \"Cryosections of 10 µm and 16 µm thickness were then obtained using a Leica CM1850 cryostat at -18 to -20 °C and mounted onto Stereo-seq chips and specially coated 50 mm × 64 mm rectangular glass slides for MERFISH+, respectively.\",\n      \"paper_location\": \"Preparation of multi-centimeter tissue sections for Stereo-seq and MERFISH+\"\n    },\n    {\n      \"entity\": \"MERFISH+\",\n      \"label\": \"Method\",\n      \"sentence\": \"Cryosections of 10 µm and 16 µm thickness were then obtained using a Leica CM1850 cryostat at -18 to -20 °C and mounted onto Stereo-seq chips and specially coated 50 mm × 64 mm rectangular glass slides for MERFISH+, respectively.\",\n      \"paper_location\": \"Preparation of multi-centimeter tissue sections for Stereo-seq and MERFISH+\"\n    },\n    {\n      \"entity\": \"Stereo-seq\",\n      \"label\": \"Method\",\n      \"sentence\": \"For Stereo-seq, tissue fixation and spatial transcriptomics procedures were conducted in accordance with the manufacturer's protocol and established methods.\",\n      \"paper_location\": \"Sample processing for Stereo-seq\"\n    },\n    {\n      \"entity\": \"methanol\",\n      \"label\": \"Chemical\",\n      \"sentence\": \"Briefly, tissue sections mounted on Stereo-seq chips were dried at 37°C (5 min), then fixed in precooled methanol (SigmaAldrich, 1060071000) for 30 mins at -20 °C.\",\n      \"paper_location\": \"Sample processing for Stereo-seq\"\n    },\n    {\n      \"entity\": \"Qubit™ ssDNA Assay Kit\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"After fixation, sections were stained with Qubit™ ssDNA Assay Kit (Invitrogen, Q10212) and incubated in the dark for 5 mins, washed with 0.1× SSC buffer with 5% RNase inhibitor (0.05 U/mL, New England Biolabs, M0314L), the nuclei-stained tissue section was imaged using the FITC channel under a 10x microscope objective of Echo Revolution microscope (Discover Echo, RRID: SCR_027699).\",\n      \"paper_location\": \"Sample processing for Stereo-seq\"\n    },\n    {\n      \"entity\": \"RNase inhibitor\",\n      \"label\": \"Protein\",\n      \"sentence\": \"After fixation, sections were stained with Qubit™ ssDNA Assay Kit (Invitrogen, Q10212) and incubated in the dark for 5 mins, washed with 0.1× SSC buffer with 5% RNase inhibitor (0.05 U/mL, New England Biolabs, M0314L), the nuclei-stained tissue section was imaged using the FITC channel under a 10x microscope objective of Echo Revolution microscope (Discover Echo, RRID: SCR_027699).\",\n      \"paper_location\": \"Sample processing for Stereo-seq\"\n    },\n    {\n      \"entity\": \"Echo Revolution microscope\",\n      \"label\": \"Instrument\",\n      \"sentence\": \"After fixation, sections were stained with Qubit™ ssDNA Assay Kit (Invitrogen, Q10212) and incubated in the dark for 5 mins, washed with 0.1× SSC buffer with 5% RNase inhibitor (0.05 U/mL, New England Biolabs, M0314L), the nuclei-stained tissue section was imaged using the FITC channel under a 10x microscope objective of Echo Revolution microscope (Discover Echo, RRID: SCR_027699).\",\n      \"paper_location\": \"Sample processing for Stereo-seq\"\n    },\n    {\n      \"entity\": \"StereoMap ImageQC\",\n      \"label\": \"Software\",\n      \"sentence\": \"The resulting TIFF file was run through StereoMap ImageQC.\",\n      \"paper_location\": \"Sample processing for Stereo-seq\"\n    },\n    {\n      \"entity\": \"cDNA\",\n      \"label\": \"NucleicAcid\",\n      \"sentence\": \"RNA captured on Stereo-seq chips from permeabilized tissue was reverse transcribed at 42°C.\",\n      \"paper_location\": \"Sample processing for Stereo-seq\"\n    },\n    {\n      \"entity\": \"Qubit dsDNA HS Assay Kit\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"After size selection, amplification, and purification, cDNA concentrations were measured with the Qubit dsDNA HS Assay Kit (Invitrogen, Q32854).\",\n      \"paper_location\": \"Sample processing for Stereo-seq\"\n    },\n    {\n      \"entity\": \"Agilent's D5000 ScreenTape assay\",\n      \"label\": \"Method\",\n      \"sentence\": \"The distribution of cDNA fragment size was also measured using Agilent's D5000 ScreenTape assay (Agilent, 5067-5588).\",\n      \"paper_location\": \"Sample processing for Stereo-seq\"\n    },\n    {\n      \"entity\": \"DNA nanoball\",\n      \"label\": \"NucleicAcid\",\n      \"sentence\": \"For library construction, 20 ng of cDNA per sample was processed with a library preparation kit (Complete Genomics, 111KL160) followed by DNA nanoball (DNB) generation.\",\n      \"paper_location\": \"Sample processing for Stereo-seq\"\n    },\n    {\n      \"entity\": \"DNB\",\n      \"label\": \"NucleicAcid\",\n      \"sentence\": \"For library construction, 20 ng of cDNA per sample was processed with a library preparation kit (Complete Genomics, 111KL160) followed by DNA nanoball (DNB) generation.\",\n      \"paper_location\": \"Sample processing for Stereo-seq\"\n    },\n    {\n      \"entity\": \"DNBSEQ™ T7 platform\",\n      \"label\": \"Instrument\",\n      \"sentence\": \"Sequencing was performed on the DNBSEQ™ T7 platform at PacGenomics (Agoura Hill, CA, USA) with PE100 flow cells and 50 bp read 1 and 100 bp read 2 (Complete Genomics, 100008555).\",\n      \"paper_location\": \"Sample processing for Stereo-seq\"\n    },\n    {\n      \"entity\": \"Gene selection\",\n      \"label\": \"Method\",\n      \"sentence\": \"1. Gene selection: First, we compiled 992 candidate genes that comprised 150 canonical cell type marker genes, 20 control genes (housekeeping genes) and 822 genes that were differentially expressed between Group level cell types in the reference Human basal ganglia snRNA-seq dataset from the Allen Brain Institute (RRID:SCR_006491) .\",\n      \"paper_location\": \"MERFISH+ design and methodology\"\n    },\n    {\n      \"entity\": \"Human basal ganglia snRNA-seq dataset\",\n      \"label\": \"DataResource\",\n      \"sentence\": \"1. Gene selection: First, we compiled 992 candidate genes that comprised 150 canonical cell type marker genes, 20 control genes (housekeeping genes) and 822 genes that were differentially expressed between Group level cell types in the reference Human basal ganglia snRNA-seq dataset from the Allen Brain Institute (RRID:SCR_006491) .\",\n      \"paper_location\": \"MERFISH+ design and methodology\"\n    },\n    {\n      \"entity\": \"Allen Brain Institute\",\n      \"label\": \"DataResource\",\n      \"sentence\": \"1. Gene selection: First, we compiled 992 candidate genes that comprised 150 canonical cell type marker genes, 20 control genes (housekeeping genes) and 822 genes that were differentially expressed between Group level cell types in the reference Human basal ganglia snRNA-seq dataset from the Allen Brain Institute (RRID:SCR_006491) .\",\n      \"paper_location\": \"MERFISH+ design and methodology\"\n    },\n    {\n      \"entity\": \"MERFISH\",\n      \"label\": \"Method\",\n      \"sentence\": \"The MERFISH library was designed using a similar pipeline described in previous studies <ref type=\\\"bibr\\\">42- 44</ref> .\",\n      \"paper_location\": \"Primary encoding probe, adaptor probe, and readout probe design\"\n    },\n    {\n      \"entity\": \"primary encoding probe\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"As therein, each primary encoding probe contains a 40-mer target sequence that is complementary to the mRNA sequence of a gene, three 20nt readout sequences unique for each gene and two 20nt PCR primers.\",\n      \"paper_location\": \"Primary encoding probe, adaptor probe, and readout probe design\"\n    },\n    {\n      \"entity\": \"mRNA\",\n      \"label\": \"NucleicAcid\",\n      \"sentence\": \"As therein, each primary encoding probe contains a 40-mer target sequence that is complementary to the mRNA sequence of a gene, three 20nt readout sequences unique for each gene and two 20nt PCR primers.\",\n      \"paper_location\": \"Primary encoding probe, adaptor probe, and readout probe design\"\n    },\n    {\n      \"entity\": \"PCR primers\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"As therein, each primary encoding probe contains a 40-mer target sequence that is complementary to the mRNA sequence of a gene, three 20nt readout sequences unique for each gene and two 20nt PCR primers.\",\n      \"paper_location\": \"Primary encoding probe, adaptor probe, and readout probe design\"\n    },\n    {\n      \"entity\": \"smFISH imaging\",\n      \"label\": \"Method\",\n      \"sentence\": \"However, we made one modification to this design by adding a 4th readout sequence in order to help us to troubleshoot MERFISH experiments through smFISH imaging.\",\n      \"paper_location\": \"Primary encoding probe, adaptor probe, and readout probe design\"\n    },\n    {\n      \"entity\": \"MERFISH codebook\",\n      \"label\": \"DataResource\",\n      \"sentence\": \"The selection of readout sequences that were concatenated to a target sequence of a gene was guided by a MERFISH codebook, which is a set of unique binary barcodes that determines in which imaging round the fluorescence signal of each gene will show up.\",\n      \"paper_location\": \"Primary encoding probe, adaptor probe, and readout probe design\"\n    },\n    {\n      \"entity\": \"Metropolis Hastings optimization algorithm\",\n      \"label\": \"Software\",\n      \"sentence\": \"Next, we used the Metropolis Hastings optimization algorithm to assign each binary word to a particular gene while minimizing the chances of molecular(signal) crowding due to shared channels when imaging genes that are co-expressed in the same cell type.\",\n      \"paper_location\": \"Primary encoding probe, adaptor probe, and readout probe design\"\n    },\n    {\n      \"entity\": \"Adaptor probes\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"Adaptor probes are associated with a particular gene and they are made of 20nt sequence reverse complementary to the readout sequence used in the primary probe of the associated gene.\",\n      \"paper_location\": \"Primary encoding probe, adaptor probe, and readout probe design\"\n    },\n    {\n      \"entity\": \"readout probes\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"The readout probes are 20nt sequences that are conjugated with a fluorescence dye at the 5' or 3' end.\",\n      \"paper_location\": \"Primary encoding probe, adaptor probe, and readout probe design\"\n    },\n    {\n      \"entity\": \"fluorescence dye\",\n      \"label\": \"Chemical\",\n      \"sentence\": \"The readout probes are 20nt sequences that are conjugated with a fluorescence dye at the 5' or 3' end.\",\n      \"paper_location\": \"Primary encoding probe, adaptor probe, and readout probe design\"\n    },\n    {\n      \"entity\": \"PCR\",\n      \"label\": \"Method\",\n      \"sentence\": \"Oligo pools were amplified through PCR with approximately 18 cycles using Phusion Plus PCR Master Mix (Thermo Fisher Scientific, F631L).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"Phusion Plus PCR Master Mix\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"Oligo pools were amplified through PCR with approximately 18 cycles using Phusion Plus PCR Master Mix (Thermo Fisher Scientific, F631L).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"RNA\",\n      \"label\": \"NucleicAcid\",\n      \"sentence\": \"RNA was generated from the purified PCR-amplified DNA through in-vitro transcription (IVT) using HiScribe T7 ARCA mRNA kit (NEB, E2060S).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"DNA\",\n      \"label\": \"NucleicAcid\",\n      \"sentence\": \"RNA was generated from the purified PCR-amplified DNA through in-vitro transcription (IVT) using HiScribe T7 ARCA mRNA kit (NEB, E2060S).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"in-vitro transcription\",\n      \"label\": \"Method\",\n      \"sentence\": \"RNA was generated from the purified PCR-amplified DNA through in-vitro transcription (IVT) using HiScribe T7 ARCA mRNA kit (NEB, E2060S).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"IVT\",\n      \"label\": \"Method\",\n      \"sentence\": \"RNA was generated from the purified PCR-amplified DNA through in-vitro transcription (IVT) using HiScribe T7 ARCA mRNA kit (NEB, E2060S).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"HiScribe T7 ARCA mRNA kit\",\n      \"label\": \"Reagent\",\n      \"sentence\": \"RNA was generated from the purified PCR-amplified DNA through in-vitro transcription (IVT) using HiScribe T7 ARCA mRNA kit (NEB, E2060S).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"single strand DNA\",\n      \"label\": \"NucleicAcid\",\n      \"sentence\": \"Next, single strand DNA (ssDNA) was generated from the IVT-derived RNA through reverse transcription (RT) using Thermo Scientific Maxima Reverse Transcriptase (Thermo Fisher Scientific, EP0743).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"ssDNA\",\n      \"label\": \"NucleicAcid\",\n      \"sentence\": \"Next, single strand DNA (ssDNA) was generated from the IVT-derived RNA through reverse transcription (RT) using Thermo Scientific Maxima Reverse Transcriptase (Thermo Fisher Scientific, EP0743).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"reverse transcription\",\n      \"label\": \"Method\",\n      \"sentence\": \"Next, single strand DNA (ssDNA) was generated from the IVT-derived RNA through reverse transcription (RT) using Thermo Scientific Maxima Reverse Transcriptase (Thermo Fisher Scientific, EP0743).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"RT\",\n      \"label\": \"Method\",\n      \"sentence\": \"Next, single strand DNA (ssDNA) was generated from the IVT-derived RNA through reverse transcription (RT) using Thermo Scientific Maxima Reverse Transcriptase (Thermo Fisher Scientific, EP0743).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"Thermo Scientific Maxima Reverse Transcriptase\",\n      \"label\": \"Protein\",\n      \"sentence\": \"Next, single strand DNA (ssDNA) was generated from the IVT-derived RNA through reverse transcription (RT) using Thermo Scientific Maxima Reverse Transcriptase (Thermo Fisher Scientific, EP0743).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"NanoDrop\",\n      \"label\": \"Instrument\",\n      \"sentence\": \"The concentration and purity of the encoding probes were determined through NanoDrop while the size of the encoding probes was verified through Agilent TapeStation (RRID:SCR_019547).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"Agilent TapeStation\",\n      \"label\": \"Instrument\",\n      \"sentence\": \"The concentration and purity of the encoding probes were determined through NanoDrop while the size of the encoding probes was verified through Agilent TapeStation (RRID:SCR_019547).\",\n      \"paper_location\": \"Encoding probe synthesis\"\n    },\n    {\n      \"entity\": \"5Alex750N/3AlexF750N\",\n      \"label\": \"Fluorophore\",\n      \"sentence\": \"5Alex750N/3AlexF750N, 5Cy5/3Cy5Sp, 5Cy3/3Cy3Sp) were purchased from IDT.\",\n      \"paper_location\": \"Adaptor and readout probe preparation\"\n    },\n    {\n      \"entity\": \"5Cy5/3Cy5Sp\",\n      \"label\": \"Fluorophore\",\n      \"sentence\": \"5Alex750N/3AlexF750N, 5Cy5/3Cy5Sp, 5Cy3/3Cy3Sp) were purchased from IDT.\",\n      \"paper_location\": \"Adaptor and readout probe preparation\"\n    },\n    {\n      \"entity\": \"5Cy3/3Cy3Sp\",\n      \"label\": \"Fluorophore\",\n      \"sentence\": \"5Alex750N/3AlexF750N, 5",
+  "source_metadata": {
+    "source_path": "evaluation/ner/multiscale_spatial_transcriptomic/2025.12.02.691876v1.full.pdf",
+    "statistics": {
+      "model": "openai/gpt-5.5",
+      "prompt_file": "/Users/pujatrivedi/Desktop/MIT/structsense/evaluation/ner/direct_llm_call/prompts/extractor_neuroscience_ner.txt",
+      "input_mode": "grobid_tei",
+      "pdf_engine": null,
+      "grobid_url": "http://localhost:8070",
+      "temperature": null,
+      "seed": null,
+      "finish_reason": "length",
+      "extracted_at": "20260702_192655",
+      "total_entities": 0,
+      "entities_by_label": {}
+    }
+  }
+}
\ No newline at end of file
diff --git a/evaluation/ner/direct_llm_call/prompts/extractor_neuroscience_ner.txt b/evaluation/ner/direct_llm_call/prompts/extractor_neuroscience_ner.txt
new file mode 100644
index 0000000..fd763de
--- /dev/null
+++ b/evaluation/ner/direct_llm_call/prompts/extractor_neuroscience_ner.txt
@@ -0,0 +1,185 @@
+You are a neuroscience-domain named-entity recognition (NER) extractor.
+You extract EXHAUSTIVELY. Recall matters more than precision.
+
+TASK
+Given neuroscience text (paper, abstract, methods section, review),
+identify EVERY mention of:
+- entities: typed neuroscience referents (proteins, regions, methods, …)
+- key_terms: salient phrases that aren't single entities but matter for
+  retrieval (paradigms, technique families, behavioral assays).
+
+EXHAUSTIVENESS — READ CAREFULLY
+- Extract EVERY occurrence. If "BDNF" appears 30 times, emit 30 entity
+  items, one per occurrence.
+- Do NOT deduplicate. Do NOT collapse repeat mentions. Do NOT emit "one
+  row per unique surface form." The post-processor handles dedup.
+- Mentions in different sentences ARE different mentions — emit all.
+- Mentions in the same sentence are different mentions — emit all.
+- Acronyms AND their expansions (e.g. "long-term potentiation (LTP)") are
+  TWO mentions sharing a label. Emit BOTH every time.
+- Symbol/full-name pairs ("Pvalb (parvalbumin)") are TWO mentions every
+  time they appear — typically Gene + Protein labels respectively.
+- Plurals, possessives, inflections ("neurons", "neuron's") are mentions
+  of the same entity — emit each occurrence with its exact surface form.
+- Methods sections in particular have very high mention density (reagents,
+  catalog numbers, instruments, protocols, statistics). Be thorough.
+- The expected count is HIGH. A typical neuroscience paper paragraph yields
+  20–60 entity mentions; a full methods section yields 200–500; a full paper
+  yields 800–2000+. If your output feels short, you are missing mentions
+  — go back and re-scan.
+
+LABEL TAXONOMY (SUGGESTED labels — prefer these, but not a closed list)
+These labels cover the common neuroscience entity types and should be your
+FIRST choice: if a mention fits one of them, use it verbatim so labels stay
+consistent across the corpus. They are guidance, NOT an exhaustive whitelist.
+When a mention is clearly an entity but none of these labels fits well, assign
+the MOST APPROPRIATE label you can — coin a concise, descriptive PascalCase
+label (e.g. `ImagingModality`, `AnatomicalAxis`) rather than forcing a poor fit
+or falling back to `Other`. Reserve `Other` for entities you genuinely cannot
+characterize. Reuse any new label consistently within a document. The judge and
+post-processor reconcile labels downstream, so a well-chosen new label is far
+more useful than a wrong one from the list.
+== Anatomy & function ==
+- BrainRegion        Macroscopic structures: hippocampus, mPFC, CA1, layer 5.
+- NeuralCircuit      Named pathways/loops: mesolimbic pathway, default mode network.
+- CorticalLayer      L1–L6 or named layers.
+- NervousSystemPart  PNS components: dorsal root ganglion, sciatic nerve.
+
+== Cells & subcellular ==
+- CellType           Neuron / glia subtypes: CA1 pyramidal neuron, microglia,
+                     parvalbumin interneuron, astrocyte.
+- CellularStructure  Subcellular components: dendritic spine, axon initial
+                     segment, postsynaptic density, mitochondrion.
+- Synapse            Synapse types or named synapses: excitatory synapse,
+                     CA3–CA1 synapse.
+
+== Molecules ==
+- Gene               Gene symbols or names: BDNF, MECP2, Fos.
+- Protein            Proteins / receptors / channels: NMDAR, tau, GluA1,
+                     Nav1.6, c-Fos.
+- Chemical           Small molecules: dopamine, glutamate, kainate, TTX.
+- Drug               Pharmacological agents with action: ketamine, propofol,
+                     muscimol.
+- Neuropeptide       Bombesin, oxytocin, NPY.
+- IonChannel         Specific channels: Kv1.2, HCN1, NaV1.6 (overrides Protein
+                     when channel-typing matters).
+- Neurotransmitter   GABA, glutamate, dopamine, serotonin.
+
+== Species & models ==
+- Species            Mus musculus, mouse, rat, zebrafish, C. elegans.
+- Strain             C57BL/6J, Sprague-Dawley, Long-Evans.
+- TransgenicLine     Pvalb-Cre, Thy1-GCaMP6f, App/PS1.
+
+== Methods & assays ==
+- Method             Techniques: patch clamp, two-photon calcium imaging,
+                     scRNA-seq, optogenetics, fMRI.
+- BehavioralAssay    Named tasks: Morris water maze, novel object recognition,
+                     fear conditioning.
+- Stimulus           Sensory or experimental stimuli: 1 kHz tone, blue light
+                     (470 nm), foot shock.
+
+== Measurements & phenomena ==
+- Measurement        Quantifiable variables: firing rate, EPSC amplitude,
+                     calcium transient, BOLD signal.
+- Phenomenon         Named effects/states: long-term potentiation (LTP),
+                     theta rhythm, sharp-wave ripple.
+- Disease            Disorders: Alzheimer's disease, autism spectrum disorder,
+                     epilepsy, schizophrenia.
+- Phenotype          Observed traits: hyperactivity, memory deficit, anxiety-
+                     like behavior.
+
+== Misc ==
+- Software           Named software/toolkits used as analytic methods.
+- Other              Clearly an entity but no label above fits.
+
+OUTPUT
+Strict JSON. No prose. No markdown fences. No comments inside JSON.
+
+The source's paper_title / doi / source_path live ONCE at the top level
+under `source_metadata`. Do NOT repeat them on every entity. With hundreds
+of mentions per paper, repeating these would inflate the JSON size 5–10x
+for zero information gain. `paper_location` (section / page) is
+per-entity because it varies.
+
+❌ WRONG — DO NOT EMIT (this is a hard rejection signal; output that
+              looks like this will be rejected as INVALID):
+{
+  "entities": [
+    {
+      "entity": "basal ganglia", "label": "BrainRegion",
+      "sentence": "...",
+      "paper_title": "Multiscale Spatial Transcriptomic Atlas",   ← WRONG
+      "doi":         "10.64898/2025.12.02.691876"                 ← WRONG
+    },
+    { "...repeated 1000 more times..." }                          ← WRONG
+  ]
+}
+
+✅ RIGHT — emit paper_title/doi ONCE at the top, never per-entity:
+{
+  "source_metadata": {                                            ← ONCE
+    "paper_title": "Multiscale Spatial Transcriptomic Atlas",
+    "doi":         "10.64898/2025.12.02.691876"
+  },
+  "entities": [
+    {
+      "entity": "basal ganglia", "label": "BrainRegion",
+      "sentence": "...",
+      "paper_location": "Introduction"      ← paper_location IS per-entity
+    },
+    { "...more entities — none with paper_title or doi..." }
+  ]
+}
+
+Schema:
+{
+  "source_metadata": {
+    "paper_title": "<title if provided in METADATA, else null>",
+    "doi":         "<doi if provided in METADATA, else null>",
+    "source_path": "<file path / url if provided, else null>"
+  },
+  "entities": [
+    {
+      "entity": "<surface form, EXACTLY as in text>",
+      "label":  "<a label from the taxonomy above, or a coined PascalCase label if none fits>",
+      "sentence": "<full sentence containing the entity>",
+      "paper_location": "<section/page/paragraph if inferable from text, else null>"
+    }
+  ],
+  "key_terms": [
+    {
+      "term": "<surface form>",
+      "sentence": "<containing sentence>",
+      "paper_location": "<section/page if inferable, else null>"
+    }
+  ]
+}
+
+RULES
+1. `entity` MUST be the exact surface form as it appears in the text.
+2. `sentence` MUST be the full sentence containing the mention, copied
+   verbatim from the input text.
+3. Emit EVERY occurrence as its own item (see "Exhaustiveness" above);
+   repeat mentions of the same surface form are separate items — do not
+   collapse them.
+4. The same surface form may appear in both entities and key_terms only if
+   it is a genuinely distinct mention; do not duplicate the identical
+   mention across both lists.
+5. Do NOT hallucinate (do not emit a mention that isn't in the text). But DO
+   include genuine in-text mentions even at ~50% label confidence — pick
+   the most likely label (preferring the suggested taxonomy, otherwise the
+   most appropriate PascalCase label you can coin); the judge handles uncertain labels later.
+6. ACRONYM HANDLING: if both expansion and acronym are in the source
+   ("hippocampus (HP)"), emit BOTH as separate entities sharing a `label`.
+   Repeat this every time the pair recurs.
+7. NEGATED MENTIONS: still emit ("no significant change in BDNF" → BDNF as Gene).
+8. LABEL DISAMBIGUATION:
+   - `Drug` overrides `Chemical` when the source describes therapeutic/
+     pharmacological use.
+   - `IonChannel` overrides `Protein` for channel proteins when the source
+     emphasizes channel function.
+   - `Phenomenon` is for named effects, not single measurements (firing rate
+     is a Measurement; LTP is a Phenomenon).
+9. If input has no entities, return {"entities": [], "key_terms": []}.
+
+If you cannot comply, output exactly: {"error": "<one-line reason>"}
diff --git a/pyproject.toml b/pyproject.toml
index 84ad189..52160dc 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -16,6 +16,9 @@ classifiers = [
   "License :: OSI Approved :: Apache Software License",
   "Operating System :: OS Independent",
 ]
+dependencies = [
+    "requests>=2.32.5",
+]
 
 [project.urls]
 homepage = "https://github.com/sensein/structsense"