Building an Autonomous AI Co-Scientist for EGFR Inhibitor Discovery

This tutorial presents an end-to-end autonomous AI co-scientist workflow for next-generation EGFR inhibitor discovery, focusing on the C797S osimertinib-resistance mutation in non-small cell lung cancer. The process starts by resolving the biological target through ChEMBL and UniProt, then mining curated EGFR IC50 bioactivity records and converting them into a clean pIC50 modeling dataset. We use RDKit to standardize molecules, remove salts, aggregate replicate measurements, compute Morgan fingerprints, extract physicochemical descriptors, and analyze scaffold diversity.

This approach enables the model to learn from chemically meaningful representations rather than raw strings. From there, we train a scaffold-split Random Forest QSAR model, evaluate its ability to generalize to unseen chemotypes, interpret potency-driving features with SHAP or model importances, and visualize influential molecular substructures. Finally, we move beyond prediction into generative design by recombining BRICS fragments from potent drug-like actives, scoring the resulting virtual analogs across potency, drug-likeness, synthesizability, novelty, and developability gates, and cross-checking the shortlisted candidates against PubChem.

The workflow begins with preparing the full scientific computing environment and installing necessary chemistry, modeling, plotting, and API dependencies. The EGFR target settings are configured, modeling constants are defined, reproducible random seeds are initialized, and helper functions for banners and robust JSON API calls are created. The ChEMBL target is resolved, UniProt context is retrieved when available, and the biological motivation around EGFR C797S resistance is framed.

Curated IC50 bioactivity measurements are mined from ChEMBL and converted into a usable pIC50 dataset. Incomplete, non-exact, or inconsistent measurements are filtered out to ensure the downstream QSAR model trains on cleaner potency values. Molecules are standardized with RDKit, salts or smaller fragments are removed, duplicate molecules are aggregated by median pIC50, and each molecule is converted into Morgan fingerprint bits plus interpretable physicochemical descriptors.

The curated chemical space is analyzed by extracting Murcko scaffolds and identifying the most common chemotype families in the dataset. Morgan fingerprints are projected with PCA to visualize how EGFR inhibitors distribute across chemical space and how potency varies. A Random Forest QSAR model is trained using a scaffold split, which helps evaluate whether the model generalizes to unseen molecular scaffolds rather than memorizing close analogs.

The trained QSAR model is interpreted by estimating which descriptors and fingerprint bits contribute most strongly to predicted potency. SHAP is used when available; otherwise, Random Forest feature importances are used to keep the workflow robust. Representative molecular substructures associated with influential ECFP bits are visualized, and generative design begins by decomposing potent drug-like parent molecules into BRICS fragments and recombining them to generate novel virtual analogs.