OPLM structure eval sets
Unsupervised contact-precision evaluation sets for the open protein language model (OPLM). Each structure is one protein chain in a PDB file whose SEQRES carries the full sequence, with residues numbered by sequence position. The model sees the full sequence (the ESM convention); residues without coordinates load as NaN rows, and contact pairs that involve them are not scored. Contacts: virtual Cβ < 8 Å, sequence separation ≥ 24.
A coevolution/ folder beside the sets holds MSAs and GREMLIN Potts contact maps for the sets
used during training, which add coevolution metrics to the structure evals.
Inventory
| folder | sets (structures) | size | tag |
|---|---|---|---|
2026-10-12/ |
casp15_16 (101), cameo (435), cameo_256 (256), pdb40 (29,429), pdb40_2048 (2,048), pdb40_256 (256) |
1.5 GiB | 2026-10-12 |
2026-10-12/coevolution/ |
MSAs and Potts maps for the 2,372 sequences of casp15_16, cameo_256, pdb40_256, pdb40_2048 |
2.1 GiB | 2026-11-10 |
The version is the snapshot date of the wwPDB fold dataset the CAMEO and PDB40 chains come
from (brineylab/oplmfold-training-data,
wwpdb/2026-10-12).
casp15_16: CASP15 and CASP16 single-domain evaluation units; full target sequence, coordinates only for the domain.cameo: CAMEO targets 2021-10 to 2022-09 (the fold head's held-out list), filtered.pdb40: one representative chain per MMseqs2 40%-identity cluster of the wwPDB snapshot. The Hub caps a directory at 10,000 files, so it ships aspdb40.tar.gz; extract it in place (see Download).*_256,*_2048: nested seed-42 subsets of chains ≤ 512 residues, for in-training evals.coevolution:msas/<sha256>.a3m.gz(ColabFoldenvMSAs: UniRef30 2302 + ColabFoldDB 2023),couplings/<sha256>.npy(float16 APC-corrected GREMLIN Potts maps) andstats/<sha256>.json(depth, kept columns, Neff), keyed by the sha256 of the full sequence.manifest.parquetgives each structure's Neff bin. Built withoplm fold eval-coevolution; itsREADME.mdhas the settings and checks.
Curation: X-ray ≤ 2.5 Å or cryo-EM ≤ 3.0 Å, at least 70% of residues modeled, at least L/5 long-range contacts. CAMEO is not held out from PDB40: all 435 CAMEO chains share a 40% cluster with a PDB40 representative.
Each version folder holds README.md (thresholds, counts by stage, source checksums),
manifest.parquet (one row per file), summary.json and sources/ (CASP downloads and
domain summary pages, the CAMEO list, the MMseqs2 FASTA and cluster table). The full methods
are in docs/DATA_MANIFEST.md of the OPLM repository. A file never changes after its tag is
created; coevolution/ was added to 2026-10-12/ later, under the tag 2026-11-10.
Download
# the sets used during training (~360 MB)
hf download brineylab/oplm-structure-evals --repo-type dataset --revision 2026-10-12 \
--include "2026-10-12/casp15_16/*" --include "2026-10-12/cameo_256/*" \
--include "2026-10-12/pdb40_256/*" --include "2026-10-12/pdb40_2048/*" \
--local-dir oplm-structure-evals
# everything; the full pdb40 is a tarball (1.1 GB) that extracts to 2026-10-12/pdb40/
hf download brineylab/oplm-structure-evals --repo-type dataset --revision 2026-10-12 \
--local-dir oplm-structure-evals
tar -xzf oplm-structure-evals/2026-10-12/pdb40.tar.gz -C oplm-structure-evals/2026-10-12
# the coevolution maps the evals read (~330 MB; add msas/ and stats/ to rebuild them)
hf download brineylab/oplm-structure-evals --repo-type dataset --revision 2026-11-10 \
--include "2026-10-12/coevolution/manifest.parquet" \
--include "2026-10-12/coevolution/couplings/*" --local-dir oplm-structure-evals
Use
data:
eval:
pdb40: {path: oplm-structure-evals/2026-10-12/pdb40_256, type: structure, every: {steps: 5_000}}
pdb40_2048: {path: oplm-structure-evals/2026-10-12/pdb40_2048, type: structure, every: {steps: 50_000}}
cameo: {path: oplm-structure-evals/2026-10-12/cameo_256, type: structure, every: {steps: 5_000}}
casp: {path: oplm-structure-evals/2026-10-12/casp15_16, type: structure, every: {steps: 5_000}}
Add coevolution: oplm-structure-evals/2026-10-12/coevolution to a set's entry for the
coevolution metrics.
A structure longer than max_position_embeddings - 2 is not scored: casp15_16 needs at
least 820, the full pdb40 1026, and the subsets 514.
Sources and licenses
cameo,pdb40: derived from the wwPDB archive, CC0 1.0. wwPDB consortium, Nucleic Acids Res. 47, D520 (2019). CAMEO target list: Haas et al., Proteins 2018.casp15_16and the CASP files insources/: target structures, sequences and domain definitions published by the Protein Structure Prediction Center for CASP15 and CASP16, redistributed unmodified (tarballs) or trimmed to one domain (PDB files). Use is subject to the Prediction Center's terms; cite the CASP15 and CASP16 assessments.- Clustering: MMseqs2 (Steinegger and Söding, Nat. Biotechnol. 35, 1026, 2017).
coevolution: MSAs from UniRef30 and ColabFoldDB through the ColabFold search (Mirdita et al., Nat. Methods 19, 679, 2022), CC BY 4.0; the maps are derived from them. Potts fit: GREMLIN (Kamisetty et al., PNAS 110, 15674, 2013).
- Downloads last month
- 293