brineylab/oplmfold-training-data

OPLMFold training data

Structural training data for OPLMFold, the structure-prediction head of the open protein language model (OPLM). Each source of structures is preprocessed once into OPLM's flat all-atom representation (oplm.fold.data.complex.MolecularComplex) and published here as a self-contained folder that oplm.fold.data.dataset.StructureDataset loads directly.

Inventory

folder	source	snapshot	entries	chains	size	license	tag
wwpdb/2026-10-12/	wwPDB mmCIF archive, protein chains	2026-10-12	254,424	1,080,220	37 GiB	CC0 1.0	wwpdb-2026-10-12

Planned sources, each in its own folder: afdb-distill/ (monomer distillation), afdb-dimers/ (the AFDB quaternary-structure release), antibody-antigen/.

Layout rules

• One folder per source and version, <source>/<version>/. A version folder never changes after its tag is created; a refreshed source is a new sibling folder, so a training config that names wwpdb/2026-10-12 always means the same bytes.
• Every version folder is self-describing: README.md, manifest.json (provenance, configuration, counts, rejections), index.parquet (one row per chain, with the split column), shards/*.parquet (one row per entry), targets/ (held-out target lists), checksums.txt.
• One tag per publish, <source>-<version>. A tag pins the whole repository as it was at that publish.
• The frontmatter license is the most permissive statement that covers every folder; the inventory table gives each source's own license.

Download

# everything
hf download brineylab/oplmfold-training-data --repo-type dataset --local-dir oplmfold-data
# one source
hf download brineylab/oplmfold-training-data --repo-type dataset --include "wwpdb/2026-10-12/*" --local-dir oplmfold-data
# only the index and manifest, to inspect before fetching the shards
hf download brineylab/oplmfold-training-data --repo-type dataset \
    --include "wwpdb/2026-10-12/index.parquet" --include "wwpdb/2026-10-12/manifest.json" --local-dir oplmfold-data

Load

pip install "oplm[fold] @ git+https://github.com/briney/oplm"

from functools import partial
from torch.utils.data import DataLoader
from oplm.fold.data.collate import collate_rows
from oplm.fold.data.dataset import StructureDataset

ds = StructureDataset("oplmfold-data/wwpdb/2026-10-12", split="train", crop_size=384, seed=0)
loader = DataLoader(ds, batch_size=8, collate_fn=partial(collate_rows, crop_size=384))

Splits in the wwPDB index: train (released on or before 2021-09-30), valid (64 seeded post-cutoff monomers), post_cutoff, cameo and casp14 (held-out evaluation targets), excluded_heldout_sequence (pre-cutoff entries sharing a chain sequence with a held-out entry), undated. See wwpdb/<version>/README.md and the OPLM docs (docs/FOLD.md §8) for the parsing rules and the column schema.

Sources and licenses

• wwPDB archive and the Chemical Component Dictionary: CC0 1.0. wwPDB consortium, Nucleic Acids Res. 47, D520 (2019).
• Held-out target lists: CAMEO (Haas et al., Proteins 2018; weekly target exports) and CASP14 (Prediction Center target list and sequences); provenance in each targets/ folder and in docs/fold/m2a/ of the OPLM repository.