A foundation model for protein-ligand affinity prediction through jointly optimizing virtual screening and hit-to-lead optimization

Abstract

The structure of the protein binding pocket governs the ligand binding affinity by providing crucial intermolecular interactions and spatial compatibility. While existing methods have leveraged these structural insights to advance affinity prediction, they often treat virtual screening and hit-to-lead optimization separately, mainly due to incompatible speed-accuracy requirements. However, these two tasks complement each other, and their integration enables broader chemical exploration while preserving focus on affinity-determining substructures. Here, we present LigUnity, a foundation model for affinity prediction that jointly embeds ligands and pockets into a shared space. In particular, LigUnity learns coarse-grained active/inactive distinction through scaffold discrimination and fine-grained pocket-specific ligand preference through pharmacophore ranking. We demonstrate the effectiveness and versatility of LigUnity on eight benchmarks across six settings. In virtual screening, LigUnity outperforms 24 methods with >50% improvement and demonstrates robust generalization to novel targets. In hit-to-lead optimization, it achieves state-of-the-art performance across split-by-time, split-by-scaffold, and split-by-unit settings, emerging as a cost-efficient alternative to free energy perturbation. We further showcase how LigUnity can be employed in an active learning framework for TYK2 to efficiently find optimal ligands. Collectively, these results establish LigUnity as a versatile foundation model for affinity prediction, offering broad applicability across the drug discovery pipeline.