olewinther/dynarepo-multichain

DynaRepo multi-chain MD dataset (curated for PLaTITO)

Multi-chain protein-protein complex MD trajectories, aggregated from the DynaRepo and DynaBench (Docking Benchmark v5.5) collections, downloaded and processed for fine-tuning PLaTITO on multimer dynamics. All trajectories run at 310 K.

Quickstart

load_example.py (in this repo) shows how to open a system and pull out sequence, chain, and trajectory data with just h5py + numpy — no PLaTITO install required:

pip install huggingface_hub h5py numpy
python load_example.py 3FXD          # fetches dynarepo_dataset_3FXD.h5 and inspects it
python load_example.py 1AK4_A000A    # works the same for accession-suffixed keys

Contents

• dynarepo_dataset_{key}.h5 — one file per system (579 files). HDF5 layout:

  {pdb_or_key}/
    resname    [L] bytes   3-letter residue names
    resid      [L] int     PDB residue numbers (reset per chain)
    chain_id   [L] int     0-based chain index per residue
    310/{replica}/ca_coords  [T, L, 3] float32, Angstrom
  

  key is the PDB code where unique across the catalog, or {pdb}_{accession} when the same PDB code backs multiple distinct DynaRepo/DynaBench projects. Every replica available for a system (usually 2-6) is included.
• embeddings/esmc_6b.pt — precomputed ESM Cambrian-6B sequence embeddings per system ({pdb_id: [L, 2560] bf16 tensor}).
• dynarepo_index.json — raw catalog metadata (accession, PDB code, chain composition) for all 930 candidate systems fetched from the API, as produced by scripts/download_dynarepo.py.
• dynarepo_full_catalog.json — per-accession composition fields (protein/nucleic-acid/lipid/carbohydrate residue counts, ligands), as produced by scripts/find_dynarepo_non_protein.py.
• dynarepo_non_protein_flagged.json — the 10 systems excluded below.
• load_example.py — standalone script demonstrating how to load a system (see Quickstart above); also in scripts/dynarepo_load_example.py in the PLaTITO repo.

Curation notes

• QC (2026-07-11, scripts/qc_dynarepo.py): every replica checked for (a) sequence-adjacent Cα-Cα bond distance, (b) frame-to-frame RMSD continuity (Kabsch-aligned), (c) minimal intra-/inter-chain Cα-Cα distance (steric clashes). 1809/1811 replicas passed cleanly.
• Non-protein exclusion (2026-07-11, scripts/find_dynarepo_non_protein.py): 10 systems whose biological assembly includes a nucleic acid, lipid, carbohydrate, or explicit ligand were removed entirely — their chain geometry is shaped by a component not represented in this CA-only protein extraction (e.g. nucleosome core particles, where the histone octamer's compact arrangement is scaffolded by DNA that isn't stored here). See dynarepo_non_protein_flagged.json for the full list and reasons.
• Systems above 1000 residues (31 systems) are present in this dataset but deliberately excluded from PLaTITO's standard training tiers as a held-out zero-shot extrapolation check — see configs/data/mixed_tiered.yaml in the PLaTITO repo.
• Duplicate removal (2026-07-15): 31 plain dynarepo_dataset_{PDB}.h5 files were removed because an equivalent {pdb}_{accession} pair already covered the same two DynaRepo/DynaBench projects for that PDB code (e.g. dynarepo_dataset_1AK4.h5 → 1AK4_A000A.h5 + 1AK4_A00PC.h5). Every other plain-named system was checked against DynaRepo's full catalog and confirmed to back exactly one project (or, where a PDB code has a second project elsewhere in DynaRepo, that second project is a different kind of simulation — e.g. a single-chain run of one component — not a duplicate of the complex already included here), so no further renames were needed.

License (resolved 2026-07-16 — repo is now PUBLIC)

Per-accession license metadata was checked via the DynaRepo REST API (LICENSE/LINKCENSE fields), split by source collection:

• DynaRepo-native (376/579 systems, 65%) — GROMACS/Amber runs: LICENSE: "Apache License 2.0" explicitly declared per accession. Permissive, redistribution permitted with attribution/notice retained.
• DynaBench-derived (203/579 systems, 35%) — NAMD runs of Docking Benchmark v5.5 complexes: no LICENSE field is populated in DynaRepo's per-accession metadata for this subset. However, the DynaBench paper's own Data Availability statement states "all the data and code used is freely available online" (links given in the manuscript), and the manuscript itself is published open access. On that basis this subset is treated as open for reuse.

This is a practical assessment, not a formal legal opinion. If you rely on this dataset under a use case with strict licensing requirements, verify terms against the original sources directly (see Source, below).

Source

This dataset aggregates two collections via the DynaRepo REST API:

• DynaRepo-native (65% of systems) — GROMACS/Amber simulations. Cite: Mokhtari O, Bignon E, Khakzad H, Karami Y. "DynaRepo: the repository of macromolecular conformational dynamics." Nucleic Acids Research 2026;54(D1):D393–D401. doi:10.1093/nar/gkaf1130
• DynaBench-derived (35% of systems) — NAMD simulations of Docking Benchmark v5.5 complexes, hosted through the same DynaRepo/MDposit API. Cite: Barlas AB, Laurent B, Karaca E, Prévost C, Sacquin-Mora S. "DynaBench: Dynamic data for the docking benchmark." Journal of Molecular Biology 2026. S0022283626000239 — the underlying complex structures/affinities in this subset trace back further to Guest JD, Vreven T, Zhou J, Moal I, Jeliazkov JR, Gray JJ, Weng Z, Pierce BG. "An expanded benchmark for antibody-antigen docking and affinity prediction reveals insights into antibody recognition determinants." Structure 2021;29(6):606-621. PMID: 33539768 (this is Docking Benchmark v5.5 itself — the structures DynaBench simulated, not the MD trajectory data).

Downloaded/processed with scripts/download_dynarepo.py in PanosAntoniadis/platito-internal (branch multi-chain). Not officially affiliated with DynaRepo or DynaBench — see their original licenses/citations for the underlying trajectory data.