PDB mmCIF Entry Index

The Protein Data Bank is the single global archive of experimentally-determined 3D structures of biological macromolecules, established in 1971 and now holding well over 230,000 entries. It stores atomic coordinates for proteins, nucleic acids, and their complexes determined by X-ray crystallography, cryo-EM, NMR, micro-electron diffraction, and integrative methods, along with the underlying experimental data (structure factors, EM maps, NMR restraints) and rich metadata covering sequence, ligands, modifications, oligomeric state, and validation reports. Every entry has a four-character PDB ID (e.g. 7PZB) and is distributed primarily in the mmCIF format, with legacy PDB-format files retained for compatibility.Operationally, the archive is jointly managed by the wwPDB consortium: RCSB PDB at Rutgers and UCSD handles deposits from the Americas and Oceania and serves as the wwPDB Archive Keeper, PDBe at EMBL-EBI handles Europe and Africa, PDBj at Osaka University handles Asia, and BMRB hosts NMR-specific data. All wwPDB sites receive synchronized weekly updates and serve the archive free of charge under CC0. Within structural biology and protein ML, the PDB is the canonical training and validation source for structure prediction (AlphaFold2/3, RoseTTAFold, Protenix, OpenFold), inverse folding (ProteinMPNN, ESM-IF), docking, MD setup, and template-based modelling, and time-cutoff splits on PDB release dates are the standard way to control for data leakage when benchmarking these models.

Splits

The split is deterministic: sha256(pdb_id) % 10 == 0 goes to test; buckets 1 through 9 go to train.

Dataset Statistics

Top experimental methods:

Top classifications:

Load With `datasets`

from datasets import load_dataset

ds = load_dataset("LiteFold/PDB")
print(ds)

row = ds["train"][0]
print(row)

Load one split directly:

from datasets import load_dataset

train = load_dataset("LiteFold/PDB", split="train")
test = load_dataset("LiteFold/PDB", split="test")

Stream rows without materializing the full table locally:

from datasets import load_dataset

streamed = load_dataset("LiteFold/PDB", split="train", streaming=True)
first_row = next(iter(streamed))

Use the mmcif_path column with hf_hub_download to fetch a structure file:

from datasets import load_dataset
from huggingface_hub import hf_hub_download

row = load_dataset("LiteFold/PDB", split="train[0]")[0]
local_path = hf_hub_download(
    repo_id="LiteFold/PDB",
    repo_type="dataset",
    filename=row["mmcif_path"],
)

Filter to X-ray structures with known resolution:

from datasets import load_dataset

train = load_dataset("LiteFold/PDB", split="train")
xray = train.filter(
    lambda row: row["experimental_method"] == "X-RAY DIFFRACTION"
    and not row["resolution_is_unknown"]
)

Columns

Source Files Used

• entries.idx
• Hub file metadata for paths under mmcif/**/*.cif.gz

The full parsed entries.idx table is also included as metadata/entries_idx.parquet. The preparation script is included at scripts/prepare_pdb_dataset.py.

Citation

@article{vallat2026rcsbpdb,
  title     = {{RCSB Protein Data Bank}: Delivering integrative structures alongside experimental structures and computed structure models},
  author    = {Vallat, Brinda and Rose, Yana and Piehl, Dennis W. and Duarte, Jose M. and Bittrich, Sebastian and Bi, Chunxiao and Segura, Joan and Zalevsky, Arthur and Sekharan, Monica R. and Webb, Benjamin M. and others},
  journal   = {Nucleic Acids Research},
  year      = {2026},
  publisher = {Oxford University Press},
  doi       = {10.1093/nar/gkaf1187}
}

📊 Structured Schema (Zero-Fabrication)

Estimated Rows: 98,824