Runtime-steered ultrashort molecular dynamics enables million-pose protein–ligand screening

Applying molecular dynamics (MD) at discovery scale is limited by the cost of evaluating thousands to millions of candidate poses. We present a docking-to-dynamics workflow that makes MD compatible with screening-scale evaluation by applying ensembles of ultrashort trajectories (0.1–5 ns) to each pose. MD-score, a ligand-RMSD stability metric, serves as both a pose-ranking score and, with a threshold, an online stopping criterion. In a conditional six-target benchmark of 88 actives and 702 inactives, MD-score achieved a pose-aware ROC-AUC of 0.868 for the top ten poses, while broad pose and velocity sampling was more effective than extending fewer trajectories. Runtime steering scaled evaluation to 1,234,620 EGFR poses while reducing simulated time to 24% of exhaustive evaluation. A prospective Marburg virus nucleoprotein screen yielded three initial minigenome-active hits among 76 tested compounds. These results position ultrashort MD as a scalable dynamic filter between docking, data-driven prediction and higher-cost validation.