Beyond the Top-Ranked Pose: Validity, Ranking Depth and Cost in Blind Docking

Learned and search-based docking are often compared on the top-ranked pose alone. On 303 benchmark complexes, each ligand was docked against its whole receptor with three configured pipelines, namely AutoDock Vina with gnina rescoring (AutoDock), the diffusion model DiffDock-L with smina optimisation (DiffDock) and EquiBind with gnina optimisation (EquiBind). No pipeline received binding-site information. A qualifying pose was PoseBusters-valid and within 2 angstroms root-mean-square deviation of the nearest deposited ligand copy, and up to thirty ranked poses were scored per complex. Raw learned poses were largely invalid, at 24.4% pooled validity for DiffDock and 2.9% for EquiBind against 99.6% for AutoDock Vina. Local optimisation repaired validity but recovered no grossly misplaced binding mode. Rank-1 recovery was 49.2% for AutoDock, 43.6% for DiffDock and 18.8% for EquiBind. The AutoDock figure rests on gnina re-ranking, because Vina's own order reached 42.9%. At 2 angstroms the leading pair was not separated at rank-1 (95% interval -1.7 to +12.8 points), whereas at 1 angstrom it was. Retrospective best-of-top-15 recovery reached 70.3% for AutoDock and 59.1% for DiffDock, a resolved gap conditional on search effort and receptor preparation. Six to seven in ten leading-pair rank-1 failures held no qualifying pose in the ranked pool. Better ranking cannot repair such failures. Charged cost per qualifying pose did not separate AutoDock and DiffDock, although per solved complex AutoDock was cheaper. A reference-free Orai1 demonstration illustrates the validity and placement screens. Variants were selected on the same benchmark, and the comparison describes these pipelines and hardware, not method classes.