A Decision-Centric Benchmark of Enrichment and Potency Ranking in CNN-Rescored Molecular Docking

Virtual screening answers two different questions with one number: enrichment, separating actives from background molecules, and potency ranking, ordering actives by measured activity. We benchmarked both for the same score on the same compounds across five CNS binding sites (SERT, GABAA benzodiazepine site, 5-HT1A, σ1, and NMDA MK-801 pore), docking ChEMBL actives, inactives, and property-matched decoys with Uni-Dock and GNINA under a held-out normalisation/evaluation split, plus an experimentally characterised negative probe excluded from all fitting. Ranking, evaluated among pChEMBL ≥ 7 actives, was negligible for every target (median Spearman ρ = 0.020) and was not recovered by data-quality filtering, multivariate null models, scaffold-aware resampling, strict assay–site matching with rebuilt decoys, or stricter curation. Enrichment proved strongly benchmark-dependent: an apparent ΔAUC gain of +0.153 against molecular weight-matched decoys decreased to ≤0.05 once the decoys were matched to seven descriptors, and the residue was not robust to compound resampling. A classifier using no structural information separated the same actives from the same decoys at ROC-AUC 0.959–0.993, and cognate ligand redocking reproduced the crystallographic pose as the top-ranked solution for one of five targets. Benchmark enrichment should not be equated with protein-specific recognition; the benchmark, the poses, and the ranking require separate validation.