Physics-Guided Active Learning for New Ligand Discovery

Structure-based drug discovery seeks to identify small molecules that bind to protein targets and modulate their function, but current workflows remain slow, costly, and data-limited. While generative models can design ligands from protein structures, they often lack physical grounding and fail to generalize beyond static training data. We propose an active learning framework that generates 3D ligands conditioned on protein pocket geometry and iteratively retrains on the top 10% of candidates, evaluated using a composite score integrating binding affinity and solvation energy. On the refined PDBBind v2019 dataset, our approach yields consistent dataset-wide improvements: median binding affinity increases from -9.3 to -9.7 kcal/mol (compared to -8.7 kcal/mol for reference ligands), median solvation free energy decreases from 1.32 to 0.94 kcal/mol, and ASKCOS scores improve from -2.17e⁴ to -2.05e⁴, reflecting enhanced synthetic feasibility. On the BRD4 benchmark, our GNINA-AMBER hybrid scoring pipeline correctly ranks a known non-binder last, validating scoring robustness. These results demonstrate that physics-guided active learning systematically enhances ligand quality, diversity, and physical plausibility, offering a scalable and generalizable approach for structure-based drug design.