Machine learning-based benchmarking of AI-powered docking for virtual screening Efficiently predicting how a small molecule will bind to a protein—a technique known as molecular docking—is critical for speeding up early stages of drug discovery. However, conventional physics-based docking can be slow and often struggles with large chemical libraries. Gu et al. compare these classical approaches with new AI-driven docking tools, investigating not just how accurately they reproduce crystal poses but also how well they enrich active compounds in virtual screening (VS). This multi-perspective benchmark helps identify potential strengths and flaws in both physics-based and AI-based methods, guiding researchers on where and how to deploy machine learning in large-scale compound discovery. The authors assembled “VSDS-vd,” a suite of datasets emphasizing real-world complexity: (1) The TrueDecoy set pairs experimental actives with genuinely assayed inactives, posing a difficult scenario for discriminating potent binders. (2) The RandomDecoy set intermixes actives with random compounds from a commercial library. (3) The MassiveDecoy set scales up to seven million decoys, mirroring industrial-scale VS. By running redocking trials, they compared CarsiDock, KarmaDock, DiffDock, and FlexPose (four AI docking tools) against four physics-based methods like Glide and LeDock, plus AI-based scoring (RTMScore, EquiScore). Pose quality was judged not just by RMSD but also by “physical plausibility” (using the PoseBusters toolkit), shining light on previously overlooked issues like protein–ligand steric clashes. Interestingly, top AI models (CarsiDock, KarmaDock) outperformed traditional methods in redocking success rate, yet physics-based tools usually offered more physically coherent poses—particularly in intermolecular checks. For actual VS tasks, physics-based Glide excelled on a challenging “TrueDecoy” scenario, whereas AI-based methods performed best on the more realistic “RandomDecoy,” suggesting diverse roles for different tools in practice. The authors found that postprocessing AI-generated poses (for example, with force fields) only modestly improved physical plausibility but did not necessarily help in ranking. Finally, they propose a hierarchical pipeline, using a swift AI approach (KarmaDock) for initial screening, then refining hits with a more accurate AI approach (CarsiDock) for final selection. This synergy can accelerate ultra-large VS campaigns while maintaining robust enrichment of promising drug leads. Paper: nature.com/articles/s42256-0…