Scrutinization on Docking Against Individually Generated Target Pockets for Each Ligand

The longstanding challenge of limited exploration in protein receptor conformational space continues to constrain the precision of molecular docking. Ensemble docking, which employs methods such as molecular dynamics simulations to generate multiple receptor conformations for docking, has improved accuracy but remains limited by incomplete sampling and an inability to fully account for ligand-induced fit. To address these limitations, we introduce the concept of individual docking, a novel approach that involves docking against receptor conformations generated individually for each ligand in the docking library. This approach has only very recently become feasible due to advances in protein structure prediction, in particular end-to-end protein-ligand complex prediction technologies exemplified by AlphaFold3 and NeuralPLexer. In this study, we performed individual docking on 27 targets from the DUD-E dataset, using a two-step protocol that integrates NeuralPLexer's inference as the receptor sampling step with subsequent physics-based docking. Our results reveal that individual docking leads to approximately 24% reduction in the enrichment factors compared with standard docking, yet it recovers different sets of active ligands. Detailed analyses of pocket and ligand conformations suggest several potential incompatibilities between deep learning-based and physics-based virtual screening tools.