Abstract
<jats:p>Large-scale prioritization of protein–ligand interactions requires dynamic information at screening scale. We present a docking-to-dynamics workflow that evaluates poses by ensembles of ultrashort molecular dynamics (MD) trajectories (0.1–5 ns) and an online ligand-RMSD stability metric, MDscore. In a conditional six-target retrospective benchmark of 88 active and 702 inactive compounds, MD-score achieved pose-aware ROC-AUC of 0.868 at top ten poses, while sampling analyses indicated that broad pose and velocity sampling was more useful than extending a small number of trajectories. Runtime steering, which terminates unstable poses and redistributes unfinished tasks, scaled the workflow to 1,234,620 docked poses from 124,245 generated compounds on Fugaku supercomputer while reducing simulated time to 24% of exhaustive evaluation. In a Marburg virus nucleoprotein screen, the workflow prioritized compounds that yielded three initial minigenome-active hits among 76 tested compounds. These results position ultrashort MD as a practical dynamic filtering layer between docking, data-driven prediction and higher-cost validation methods.</jats:p>