Robust Multi-Objective Reinforcement Learning Improves Joint Predicted-Potency and Drug-Likeness Yield in EGFR Molecule Generation

Abstract

Optimizing a molecular generator against a single activity-model estimate can reward structures for which that estimate is unreliable. We extended our published ChemBERTa–reinforcement-learning (RL) workflow for human epidermal growth factor receptor (EGFR) with scaffold-separated activity modeling, empirically calibrated lower scores, bounded chemistry rewards, and matched-seed evaluation. From ChEMBL 37, we curated 6,476 unique structures with Reference/unspecified biochemical EGFR IC50 measurements; 4,491, 680, and 1,305 structures were assigned to scaffold-disjoint training, calibration, and test sets. The random-forest reward model achieved a test mean absolute error of 0.610 pIC50 units and R2 = 0.615. Uniform or potency-enriched masked-language-model continuation was crossed with three RL rewards, with three seeds and 600 generation attempts per combination. The locked LCB-versus-mean-only comparison changed GNN potency-screen yield by +0.28 percentage points (95% descriptive t interval,-1.10 to +1.66) across three seed-cluster means. In the pipeline-level comparison with the published-form baseline, Robust-LCB increased the GNN joint-screen yield in all six matched blocks; averaged across the two MLM arms, the yield rose from 0.72% to 1.22%. Two structures passed the full RF–GNN consensus funnel. These results support a pipeline-level increase in joint predicted-potency and drug-like yield, while the incremental contribution of the lower-score term alone remains uncertain. These comparisons describe computational variability; activity estimates for generated molecules require biochemical testing.