ADME-Classifier

Using machine-learning algorithms and genomics features to build binary classifiers that distinguish human ADME (Absorption, Distribution, Metabolism, and Excretion) gene variants from non-ADME variants.

---

Overview

ADME genes govern how drugs are absorbed, distributed, metabolized, and excreted in the human body. This project frames variant classification as a supervised binary classification problem: given a set of genomic and conservation features for a single nucleotide variant (SNV), predict whether it falls in an ADME gene.

The main experimental notebook is Different_ML_models.ipynb. It covers the full ML workflow from data loading through model evaluation with chromosome-based cross-validation.

---

Features / Inputs

Each variant is represented by the following feature groups:

#	Feature group	Description
1	phastCons scores	6 conservation scores from the phastCons track
2	phyloP scores	6 scores from the phyloP conservation track
3	vepDistance	Distance to the nearest annotated transcript (VEP)
4	Base-change dummies	One-hot encoded nucleotide substitution type (e.g. A→C, A→G, …)
5	VEP / Ensembl consequence	Functional consequence annotation (e.g. missense, synonymous)
6	SIFT / PolyPhen scores	Predicted functional impact of the amino-acid change
7	Population frequencies	Allele frequency information

Labels are stored in a Label column; chromosome, reference allele, and alternate allele columns are excluded from the feature matrix.

---

Workflow

Load CSV data
    │
    ▼
One-hot encode categorical columns (pd.get_dummies)
    │
    ▼
Chromosome-based split preparation
  ├─ Leave-One-Chromosome-Out (LOCO) folds
  └─ Simple train / validation / test split (25 % test)
    │
    ▼
StandardScaler normalisation
    │
    ▼
Model training & evaluation
  ├─ SGDClassifier  (linear baseline)
  ├─ XGBClassifier  (default + Bayesian hyperparameter search)
  └─ GradientBoostingClassifier
    │
    ▼
Probability thresholding
  └─ Convert predict_proba output to labels using a confidence threshold;
     variants below the threshold are omitted and the omitted % is reported

---

Models Tried

Model	Notes
SGDClassifier	Linear baseline; trained per LOCO fold and with K-fold CV
XGBClassifier	Default hyperparameters; Bayesian optimisation (scikit-optimize BayesSearchCV); approx tree method variant
GradientBoostingClassifier	Default hyperparameters; Bayesian optimisation

---

Evaluation Strategy

Two evaluation schemes are used:

• Leave-One-Chromosome-Out (LOCO): In each fold, all variants on one chromosome form the test set and the rest are used for training. This tests generalization to chromosomes unseen during training.
• Simple split: A standard random 75 / 25 train-test split with a nested validation set for hyperparameter tuning.

Metrics reported:

• Accuracy – average per-fold accuracy and whole-dataset accuracy
• ROC AUC – area under the receiver operating characteristic curve
• Omitted % – fraction of variants whose predicted probabilities fall below the chosen confidence threshold (probability thresholding experiments only)

Example results recorded in notebook comments (LOCO, XGBoost, no optimisation):

Average Accuracy : 61.20 %
Whole Accuracy   : 62.57 %
AUC Score        : 0.6163

With probability thresholding (threshold = 0.8):

Whole Accuracy   : 68.12 %   Omitted: 31.44 %
AUC Score        : 0.6691

These figures are illustrative outputs from the notebook and may vary with different random seeds or data versions.

---

Repository Layout

ADME-Classifier/
├── Different_ML_models.ipynb          # Main experiment notebook
├── Different_ML_models_Box_plot.ipynb # Box-plot visualisations of results
├── Different_ML_models_Thresholding_results_code.ipynb  # Thresholding analysis
├── Baseline_XGBoost_Model_LeaveOneChrom_OtherTesting.ipynb
├── XGBoost_Pars_Optimised_for_4_training_feature_groups.ipynb
├── *_Features_Extraction*.ipynb       # Feature extraction pipelines
├── Dataset_Expansion*.ipynb           # Dataset expansion experiments
├── Extracted Features/                # Pre-extracted feature files
├── dataS/                             # Supporting data directory
└── thresholding results (graphs)/     # Saved threshold sweep plots

---

Requirements / Setup

The notebooks are designed to run in Google Colab with Google Drive mounted. Key dependencies:

numpy
pandas
scikit-learn
xgboost
scikit-optimize   # pip install scikit-optimize
numba             # pip install numba
matplotlib

Install any missing packages at the top of the notebook or with:

pip install scikit-optimize numba xgboost

---

Usage / Reproducibility

1. Open the notebook in Colab using the badge at the top of this file, or clone the repo and open locally with Jupyter.
2. Mount Google Drive and set the base variable to the folder that contains your input CSV (e.g. Odata_fzero.csv).
3. Run all cells in order:
   • Data loading and one-hot encoding
   • Chromosome split preparation
   • Feature scaling
   • Model training and LOCO evaluation
   • (Optional) Probability thresholding sweep
4. Results (accuracy, AUC, omitted %) are printed inline and can be collected for the box-plot notebook.

---

Contributing

Pull requests and issues are welcome. Please open an issue first to discuss any significant changes.