caijihuize/pdb-3di-dataset

Experimental PDB AA–3Di Dataset

Paired amino acid (AA) and Foldseek 3Di sequences derived from experimentally determined protein structures in the wwPDB PDB Core Archive. Version pdb_v1 was built from a frozen RCSB Search API inventory of 255,119 released experimental PDB entries containing protein polymers. This is the source inventory, before method, quality and chain filtering.

Split	Records	Evaluation mmCIF files
Train	134,637	Not included
Validation	1,000	1,000
Test	1,000	1,000

Each Parquet row contains id, sequence_3di, and sequence_aa. Ordinary record IDs have the form <PDB_ID>_<chain ID>; multi-model entries use <PDB_ID>_MODEL_<model number>_<chain ID>. ID lists in ids/ follow Parquet row order. The corresponding *_structure_ids.tsv files give the row index, split, stable ID, PDB ID, model number, author chain ID and internal Foldseek database key. The Foldseek key is build-specific; use the stable ID externally. The full deposited mmCIF entries for validation and test are in structures/. Other chains may be present in each coordinate file.

from datasets import load_dataset

ds = load_dataset("caijihuize/pdb-3di-dataset")
row = ds["test"][0]
print(row["id"], row["sequence_3di"], row["sequence_aa"])

Method

The coordinate source is the wwPDB PDB Core Archive, accessed through RCSB. The primary version uses X-ray entries with reported resolution at most 3.0 Å and electron microscopy entries with reported resolution at most 4.0 Å. Only Foldseek-observed protein residues are included in paired sequences; unresolved residues of a deposited polymer are not added as AA-only positions. NMR and integrative/hybrid structures are outside the primary version.

Protein chains must have at least 30 observed residues, at most 20% unknown AA, matching AA/3Di lengths, and valid Foldseek 3Di symbols. MMseqs2 clustering uses 30% identity and 80% bidirectional coverage. Chains from the same PDB entry and exact AA duplicates stay in the same split group. Validation and test each select one representative from 1,000 groups, with evaluation length 30–510 residues. A final all-member sequence search and deterministic repair found no cross-split alignment meeting the published 30% identity and 80% bidirectional coverage threshold. Foldseek structural neighbors are reported in reports/ and do not control split membership.

Reproducibility and limitations

The inventory query and date, per-file checksums, method and chain metadata, group membership, exclusions, tool versions, audit summaries and validation results are in provenance/ and reports/. Verify the release with sha256sum -c checksums/SHA256SUMS after downloading all files.

PDB entries may represent only a domain or an engineered construct and may contain missing coordinates. Experimental structures have method-dependent quality and are not necessarily full-length natural proteins. Sequence isolation under the reported search settings does not establish fold novelty or absence of exposure during pretraining. The raw PDB archive and training structure coordinates are not mirrored here.

PDB archive data files are available under CC0 1.0. Cite the original PDB entries and their structure publications where possible, as well as the wwPDB/RCSB archive and Foldseek method. Source and code: https://www.wwpdb.org/, https://www.rcsb.org/, https://github.com/caijihuize/pdb-3di-dataset.