Protein Structure · LiteFold
PDB
Tabular index and mirror of 98,824 Protein Data Bank mmCIF macromolecular structures for structural machine learning.
What PDB is
The Protein Data Bank (PDB) is the central global repository of experimentally determined three-dimensional structures of biological macromolecules, including proteins, nucleic acids, and multi-component complexes. Maintained by the Worldwide Protein Data Bank (wwPDB) consortium—comprising RCSB PDB, PDBe, PDBj, and BMRB—the archive contains atomic coordinates alongside experimental parameters and functional metadata.
LiteFold/PDB provides a tabular index and file collection mirroring 98,824 macromolecular Crystallographic Information File (mmCIF) entries from the PDB. It couples individual coordinate files with tabular metadata derived from the PDB index, standardising fields such as accession dates, functional classifications, source organisms, experimental methods, and structural resolutions.
Within computational biology, this packaged dataset serves as primary data for training and evaluating machine learning systems, including macromolecular structure prediction algorithms, inverse protein folding architectures, molecular docking suites, and molecular dynamics simulation pipelines.
How it was made
The dataset is generated by pairing mirrored, gzipped mmCIF coordinate files with metadata extracted from the PDB entries index (`entries.idx`). Accession dates are parsed into ISO format, and raw resolution annotations are converted into numerical values in Ångströms alongside indicator flags for entries lacking numeric resolutions. The complete parsed index table containing 252,816 entries is also provided in Parquet format.
Entries are divided into deterministic partitions using a cryptographic hash: `sha256(pdb_id) % 10`. Bucket 0 forms the test split (9,951 structures), whilst buckets 1 through 9 constitute the training split (88,873 structures).
Key facts
- Mirrored mmCIF structures
- 98,824
- Compressed archive size
- 31.08 GB
- Full entries index rows
- 252,816
- Median resolution
- 2.10 Å
- Mean resolution
- 2.33 Å
- Dominant experimental method
- X-ray diffraction (82,380 entries)
- Dataset split sizes
- 88,873 train; 9,951 test
What you can use it for
- Protein structure prediction training
- Supplying experimentally derived 3D coordinates and metadata to train or fine-tune models such as AlphaFold, RoseTTAFold, Protenix, and OpenFold.
- Inverse folding sequence generation
- Providing backbone atomic coordinates across diverse protein families to train sequence design tools such as ProteinMPNN and ESM-IF.
- Molecular docking and binding analysis
- Filtering experimental macromolecular structures by resolution and method to benchmark receptor-ligand and complex docking workflows.
- Molecular dynamics system configuration
- Screening and downloading curated mmCIF files meeting experimental resolution thresholds for molecular dynamics preparation.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from LiteFold/PDB at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 88,873 rows of default/train
accession_date
string_text- min
- 8
- median
- 8
- mean
- 8
- max
- 8
accession_date_iso
datetimeauthors
string_text- min
- 6
- median
- 47
- mean
- 63.79
- max
- 999
classification
string_text- min
- 0
- median
- 13
- mean
- 14.55
- max
- 67
experimental_method
string_label- X-RAY DIFFRACTION83%
- ELECTRON MICROSCOPY12%
- SOLUTION NMR5%
has_entries_idx_metadata
bool- True100%
mmcif_blob_id
string_text- min
- 40
- median
- 40
- mean
- 40
- max
- 40
mmcif_file_size_bytes
int- min
- 7,323
- median
- 158k
- mean
- 316k
- max
- 102M
mmcif_path
string_text- min
- 20
- median
- 20
- mean
- 20
- max
- 20
pdb_id
string_text- min
- 4
- median
- 4
- mean
- 4
- max
- 4
pdb_url
string_text- min
- 35
- median
- 35
- mean
- 35
- max
- 35
raw_resolution
string_text- min
- 0
- median
- 3
- mean
- 3.48
- max
- 11
rcsb_download_url
string_text- min
- 43
- median
- 43
- mean
- 43
- max
- 43
resolution_angstrom
float- min
- 0
- median
- 2.10
- mean
- 2.33
- max
- 50
4.9% empty
resolution_is_unknown
bool- False95%
- True5%
source_organism
string_text- min
- 0
- median
- 20
- mean
- 23.31
- max
- 798
split_bucket
int- min
- 1
- median
- 5
- mean
- 5.00
- max
- 9
title
string_text- min
- 3
- median
- 75
- mean
- 81.25
- max
- 390
Strengths and limits
Good at
- Directly pairs tabular metadata with mirrored mmCIF file paths and external RCSB download links.
- Standardises experimental resolution into numerical Ångström values with explicit flags for missing or non-numeric measurements.
- Provides a deterministic, reproducible hash partition across all mirrored entries.
Watch out for
- Contains 98,824 mirrored mmCIF files, covering only a portion of the 252,816 entries catalogued in the full index.
- The deterministic identifier hash split does not group entries by sequence homology or chronological cut-offs, which are standard for avoiding data leakage in structural benchmark models.
- A subset of 4,827 entries lack numerical resolution metrics, which is standard for techniques such as solution NMR.
Questions
Can I look inside PDB without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is PDB under?
Its authors declare cc0-1.0. Read the terms on the original page before reusing the data.
What primary structural file format is used in this dataset?
Structures are provided as gzipped macromolecular Crystallographic Information Files (mmCIF), accessible via repository file paths and direct RCSB URLs.
How are entries without numeric resolution represented?
Entries where resolution is not numerically defined (such as solution NMR structures or entries marked 'NOT') have null values in the numerical resolution column and are flagged as true in the unknown resolution column.
How are the training and test splits partitioned?
Splits are determined by computing `sha256(pdb_id) % 10`; entries evaluating to 0 are assigned to the test split (9,951 entries), and all other values are placed in the training split (88,873 entries).
Which experimental techniques are represented in the dataset?
The indexed structures are predominantly determined by X-ray diffraction (82,380), electron microscopy (11,433), and solution NMR (4,707), alongside smaller sets from electron crystallography (101) and combined X-ray and neutron diffraction (50).
How to cite it
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 (2026). RCSB Protein Data Bank: Delivering integrative structures alongside experimental structures and computed structure models. Nucleic Acids Research. doi: 10.1093/nar/gkaf1187.
Written from PDB's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.