Six-target retrospective benchmark docking poses and MD-scores for runtime-steered ultrashort molecular dynamics

This dataset accompanies the manuscript “Runtime-steered ultrashort molecular dynamics enables million-pose protein–ligand screening.” The release contains docking-pose and processed MD-score data for a six-target retrospective benchmark covering EGFR, PTPN1, ERK2, PPARG, BACE1, and p38. The joint pose-and-activity benchmark contains 7,695 docking poses, and the complementary pose-only benchmark contains 4,117 labeled poses. After removal of overlapping poses, the release comprises 10,096 unique benchmark poses represented by 1,035 original multi-model AutoDock Vina PDBQT files. The release includes: original multi-model docking-pose PDBQT files; pose-level AutoDock Vina docking scores; pose-level MD-scores calculated from 5-ns simulations with five independent initial velocities; protein, compound, and pose-level metadata; joint and pose-only benchmark membership tables; file manifests, data dictionaries, and SHA-256 checksums. The exact benchmark poses are identified by target, category, file key, and MODEL number in metadata/benchmark_pose_information.tsv. Fourteen original PDBQT files contain a total of 20 additional MODEL records that were not used in the 10,096-pose benchmark; these records were retained to preserve the original docking outputs. MD trajectory files (XTC/GRO) are not included because of their large total data size. They are retained by the authors and may be made available upon reasonable request, subject to the authors’ data-sharing policy. The generated EGFR-focused million-pose library, the prospective MARV NP screening dataset, and experimental compound-library structures are not included in this release.