London, N.; Shamir, Y.

Recent years have seen an explosion in the prominence of covalent inhibitors as research and therapeutic tools. However, a lag in application of computational methods for covalent docking slows progress in this field. AI models such as AlphaFold3 have shown accuracy in ligand pose prediction but were never assessed for virtual screening. We show that AlphaFold3 reaches near-perfect classification (average AUC=98.3%) of covalent active binders over property-matched decoys, dramatically outperforming classical covalent docking tools. We identify a predicted metric that allows to reliably assign a probability of binding and demonstrate it also improves non-covalent virtual screening.