AI Co-Scientist Workflow Accelerates EGFR Inhibitor Discovery

An end-to-end autonomous AI co-scientist workflow has been developed to accelerate the discovery of next-generation EGFR inhibitors, specifically targeting the C797S-osimertinib-resistance mutation in non-small cell lung cancer. The process begins by identifying the biological target using ChEMBL and UniProt, followed by mining curated EGFR IC50 bioactivity records and converting them into a clean pIC50 modeling dataset. To ensure chemically meaningful representations, RDKit is employed to standardize molecules, remove salts, aggregate replicate measurements, compute Morgan fingerprints, and extract physicochemical descriptors, while also analyzing scaffold diversity. This approach prevents the model from learning from raw string representations.

The workflow then proceeds to train a scaffold-split Random Forest Quantitative Structure-Activity Relationship (QSAR) model. The model's ability to generalize to unseen chemotypes is rigorously evaluated. Potency-driving features are interpreted using SHAP (SHapley Additive exPlanations) or model importances, and influential molecular substructures are visualized. This interpretation phase is crucial for understanding the underlying chemical drivers of drug efficacy.

Moving beyond mere prediction, the AI co-scientist engages in generative design. It recombines BRICS (Bond Reconnection In Complex Structures) fragments from potent drug-like active compounds. The resulting virtual analogs are then scored against multiple gates, including potency, drug-likeness, synthesizability, novelty, and developability. Finally, the shortlisted candidates are cross-checked against data available in PubChem to validate their potential and novelty. This comprehensive approach aims to streamline the drug discovery pipeline by integrating prediction, interpretation, and generation within a single autonomous system.