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

Protein-ligand binding affinity plays an important role in drug discovery, especially during virtual screening and hit-to-lead optimization. Computational chemistry and machine learning methods have been developed to investigate these tasks. Despite the encouraging performance, virtual screening and hit-to-lead optimization are often studied separately by existing methods, partially because they are performed sequentially in the existing drug discovery pipeline, thereby overlooking their interdependency and complementarity. To address this problem, we propose LigUnity, a foundation model for protein-ligand binding prediction by jointly optimizing virtual screening and hit-to-lead optimization. In particular, LigUnity learns coarse-grained active/inactive distinction for virtual screening, and fine-grained pocket-specific ligand preference for hit-to-lead optimization. We demonstrate the effectiveness and versatility of LigUnity on eight benchmarks across both tasks. In virtual screening, LigUnity outperforms 24 competing methods with more than 50% improvement on the DUD-E and Dekois 2.0 benchmarks, and shows robust generalization to novel proteins. In hit-to-lead optimization, LigUnity achieves the best performance on split-by-time, split-by-scaffold, and split-by-unit settings, further demonstrating its potential as a cost-effective alternative to free energy perturbation (FEP) calculations. We further showcase how LigUnity can be employed in an active learning framework to efficiently identify active ligands for TYK2, a therapeutic target for autoimmune diseases, yielding over 40% improved prediction performance. Collectively, these comprehensive results establish LigUnity as a versatile foundation model for both virtual screening and hit-to-lead optimization, offering broad applicability across the drug discovery pipeline through accurate protein-ligand affinity predictions.