EGFR (NSCLC) pIC50 Predictor
ExtraTrees + tuned GAT — best model (test RMSE 0.644 · external FDA-15 RMSE 0.680)
Model
ExtraTrees + tuned GAT (graph attention) combo — the best model of the EGFR NSCLC project.
- Features: Morgan fingerprint + physchem (1,030) + tuned 2-layer GAT graph embedding (128) = 1,158
- Heads: tuned ExtraTrees regressor (pIC50) and classifier (potency class)
- Classes: potent pIC50 ≥ 6 · moderate 5–6 · not potent < 5
Performance (clean v2 dataset — ChEMBL + BindingDB + PubChem, parent-level leakage audit 0/15)
| Split | RMSE | MAE | R² | Accuracy | F1 macro | MCC |
|---|---|---|---|---|---|---|
| Validation (n=1,264) | 0.661 | — | — | — | — | — |
| Test (n=1,264) | 0.644 | 0.460 | 0.795 | 0.880 | 0.771 | 0.686 |
| External FDA-15 (n=14) | 0.680 | 0.520 | — | — | — | — |
External set = 15 approved EGFR drugs (FDA + other regulators), fully held out of training. Predictions from this app reproduce the verified external grid exactly (e.g. gefitinib 7.526, erlotinib 7.818).