Ranking metrics conceal pose errors in blind docking with AutoDock Vina, a cross-target benchmark

Abstract

Molecular docking often relies on a blind search box spanning the whole protein when the binding site is unknown or a single protocol is applied across many targets, with results accepted based on docking score or enrichment of known actives. Existing accuracy assessments re-dock ligands with known experimental poses and effectively given sites, so they fail to establish whether these ranking metrics actually track pose correctness under blind search. We tested this directly with AutoDock Vina across eleven structurally diverse targets, each with thirty actives and roughly 240 property-matched decoys, docked twice under identical settings: once with a site-directed box and once with a blind, whole-protein box. Alongside standard virtual-screening metrics, we measured how far each top-ranked pose fell from the known site. Ranking was only modestly worse under blind docking (mean ROC-AUC 0.62 vs. 0.69, p=0.032; enrichment at 5% 2.24 vs. 3.47, p=0.012), but placement collapsed, from 96% to 48% of actives within 8 Å of the site (95% CI for the difference 0.31–0.63, p=0.002). The two quantities were unrelated across targets (Pearson r=-0.03): good blind-docking ranking told us nothing about pose correctness, and two targets ranked at least as well under blind docking while placing fewer than one in five actives correctly. Pose-to-site distance tracked heavy-atom RMSD closely (r=0.98, n=2971 ligand pairs), and doubling default exhaustiveness left six of eight gross misplacements unresolved. We conclude that score- and enrichment-based validation cannot detect blind-docking placement failure, and propose a simple check to catch it.