Interactive, sanitized demos exploring where computational models actually change drug-design decisions — and where they don't.
Each demo runs a real model's predictions against experimental data from a live medicinal-chemistry program, then builds visualizations that test — honestly — how much the model's accuracy translates into useful decisions. Compounds appear with opaque IDs only; no structures, assay names, or chemical matter are published.
Matched-pair / R-group analysis of an internal IC50 model vs. experiment.
The model is excellent at absolute potency (r ≈ 0.91) — but design is about differences, and there the story turns. A swap's error is mostly noise that cancels on averaging; the only real signal is a small, reproducible per-fragment bias — and it's recalibratable. The catch: you can't shortcut the matched pairs. "This fragment looks bad in the absolute plot" predicts the swap bias at r ≈ 0 — you have to deconfound the context (r ≈ 0.6) or compare matched pairs directly. The Reality-check view shows this; a value-of-information panel turns it forward into what to make next — recalibrate the known biases for free, measure only the unresolved ones.
Project timeline + policy simulator for an internal RBFE (relative binding free-energy) model.
The model is accurate within a conformation but a coin-flip on the conformational-switch edges that unlock the best chemistry — the program's top compounds are favorable molecules sitting one false-negative crossing past the model's rejection boundary. Run a model-trusted policy (make everything it scores favorable, respecting that you can't score a compound without its reference) and watch it cap out; add expert overrides (wave through the few crucial compounds the model would skip) or an explore budget, and read the recovery off a live dial. The strategy-efficiency panel is the punchline: RBFE enriches the potent bulk (pIC50 ≥ 8) but adds little over blind exploration for the best-in-class (≥ 9) — because the best compounds are exactly the ones it said skip.
The thread: an accurate model is not automatically a useful one. These tools are how I tell the difference — benchmark against experiment, separate model limits from data noise, and find where human judgment still earns its keep.
Static HTML (Plotly + inline SVG, no build step). Open via GitHub Pages, or locally run python -m http.server inside an app folder and open index.html.
Committed files contain only opaque compound IDs and numeric values. Structures, real IDs, assay/vendor names, and the build scripts are git-ignored and never published; the viewers degrade gracefully when the private data is absent.