EGFR L858R,T790M vs strict WT pairwise Δ-modeling and selectivity-reward REINVENT

Paired ChEMBL_37 bioactivity archive for EGFR (CHEMBL203) comparing L858R,T790M against strict WT same-document measurements, with LightGBM/Chemprop Δ-prediction and REINVENT 4.8.24 selectivity-reward matrices. v1.1.0 makes HC2 (Binding–Binding + matched BAO; n=362) the primary cohort and archives HC2 formal REINVENT, orthogonal Chemprop HV, HC nest Chemprop/LightGBM metrics, multi-seed scaffold LightGBM tables, and novelty/Tanimoto vs HC2 train/Full. Full (n=748) and HC1 remain sensitivity. Primary scaffold split is deterministic Murcko greedy (no random split seed); seed 42 is for model training / Chemprop / REINVENT / bootstrap only. Dual REINVENT arms and docking case studies were not executed.