DrugAutoML

DrugAutoML is an end-to-end Automated Machine Learning (AutoML) pipeline for bioactivity prediction in drug discovery.
It automates every stage — from reading raw datasets to generating predictions for new molecules — and produces both high-performance models and explainable outputs.

🚀 Features

• Data Preprocessing
  Reads raw datasets, cleans and standardizes SMILES, removes invalid molecules, and labels compounds as active or inactive based on pChEMBL cutoffs or existing binary labels.

• Molecular Featurization
  Generates ECFP (Extended-Connectivity Fingerprints) using RDKit with customizable radius, bit size, and count-based features.

• Data Splitting
  Splits data into training and testing sets using:

  • Scaffold Split (structure-aware)
  • Stratified Random Split (class-proportion preserving)

• Model Selection
  Hyperparameter optimization with Optuna for:

  • Random Forest, Extra Trees, Logistic Regression, Linear SVC, XGBoost, LightGBM
    Uses repeated stratified k-fold CV and produces a ranked leaderboard.

• Model Finalization
  Trains the best model, applies probability calibration, selects optimal classification threshold, evaluates on the test set, and saves the model.

• Explainability

  • SHAP global importance plots (beeswarm, bar, signed bar)
  • Bit Gallery visualizations: highlights ECFP bits in test molecules that strongly influence predictions.

• Prediction on New Data
  Scores unlabeled or labeled molecules, outputs probabilities and predictions, and computes metrics if labels are available.