Physics-Based and Generative Ensemble Analysis of Macrocyclic Peptide Recognition at the pre-miR-21 Dicer-Processing Site

Data, scripts supporting a comparison of explicit-solvent molecular dynamics (MD) and BioEmu-2 generative ensembles for the 14-residue macrocyclic peptide L50 bound to, and free of, pre-miR-21 (PDB 5UZZ). Contents holo/, apo/: one 500 ns OpenMM trajectory per state (10,000 frames saved every 50 ps; bound system 27,539 atoms, free system 5,523 atoms), with topology and coordinate files, serialized system and integrator, run metadata, the state log and per-frame contact metrics. The simulations started from NMR model 1 using AMBER ff14SB (peptide), OL3 (RNA) and TIP3P-FB water at 310 K and 1 bar. Seeds: 101 (bound), 201 (free). bioemu2_holo/, bioemu2_apo/, bioemu2_query contains bioemu2_inputs/, bioemu2_logs/, bioemu2_analysis/: 32 BioEmu-2 structures per state (bioemu2-preview checkpoint, equilibrium mode, edm_gd sampler, query-only A3M alignment), raw sample batches, system definitions, the sampling log and analysis outputs. Structures contain heavy atoms only. derived_tables/: data underlying the figures and tables: backbone dispersion, contact distances, ensemble-size sensitivity, and per-record NOE-derived restraint scoring for MD and BioEmu-2 (550 of 699 deposited proton-pair records mapped and scored). scripts/: OpenMM preparation and production script, BioEmu-2 runner and sampling script, and analysis script, with placeholder paths (see scripts/README.md), and the conda environment file. documentation/: evaluation criteria fixed before the comparison. inputs/: PDB 5UZZ and the wwPDB NMR restraint files (public, included for convenience). Notes for reuse. Each state has a single MD trajectory. BioEmu-2 samples are independent structures without a time axis, and no physicality filtering was applied. The seed of the BioEmu-2 query-only run was not recorded. Hydrogens used in the restraint comparison were reconstructed from MD-reference geometry (method in documentation/), and the hydrogenated structures themselves are not included. Wall times in the associated paper were measured on different GPUs and are not hardware-normalized.