pdb-captions-2026 / README.md
matteopariset's picture
Upload README.md with huggingface_hub
d321fa6 verified
|
Raw History Blame Contribute Delete
39.1 kB
metadata
license: cc-by-4.0
pretty_name: Protplex PDB Captions (2026 Slice)
language:
  - en
size_categories:
  - 10K<n<100K
task_categories:
  - text-retrieval
  - feature-extraction
tags:
  - biology
  - protein-structure
  - structural-biology
  - pdb
  - chemistry
configs:
  - config_name: entries
    default: true
    data_files:
      - split: train
        path: data/entries.parquet
  - config_name: pockets
    data_files:
      - split: train
        path: data/pockets.parquet
  - config_name: interfaces
    data_files:
      - split: train
        path: data/interfaces.parquet
  - config_name: ligands
    data_files:
      - split: train
        path: data/ligands.parquet
  - config_name: provenance
    data_files:
      - split: train
        path: data/provenance.parquet

Protplex PDB Captions (2026 Slice)

Text captions of PDB entries' biology and structure, paired with the structural measurements they are built from (secondary structure, solvent accessibility, assembly interfaces, pockets, ligand interactions), the annotations behind them and the provenance of every value. This is the 2026 slice of the PDB as processed by Protplex: entries first released between 1 September 2025 and 30 September 2026, plus earlier entries that were new to Protplex's corpus in the same run.

Version 0.1.0 · released 2026-11-12 · 19,567 PDB entries · licence CC BY 4.0

This first release covers the PDB entries that Protplex has processed with its corrected ("v2") structural pipeline. It is published so researchers can try the data and tell us what it is useful for; feedback is welcome through a discussion on the dataset page.

Authors

Contents

Table (configuration) Grain Rows Contents
entries one row per PDB entry 19,567 Texts, annotations and entry-level structural measurements.
pockets one row per predicted pocket 196,584 Predicted pockets: the top 3 fpocket pockets and up to 10 P2Rank pockets per entry.
interfaces one row per chain pair in contact 72,256 Chain-pair interfaces of biological assembly 1.
ligands one row per PLIP binding site 120,784 Bound ligands and their measured interactions (PLIP 2.3.1).
provenance one row per entry and processing step 176,099 How and from what each value was computed.

All tables share the key pdb_id. Each table is a configuration of this dataset with a single split, train; the split is not a train/test partition.

Quick start

from datasets import load_dataset

entries = load_dataset("protplex/pdb-captions-2026", "entries", split="train")        # the default configuration
pockets = load_dataset("protplex/pdb-captions-2026", "pockets", split="train")
print(entries[0]["pdb_id"], entries[0]["structural_text"])

# pin this release
entries = load_dataset("protplex/pdb-captions-2026", "entries", split="train", revision="v0.1.0")

With pandas or DuckDB, read the Parquet files directly:

import pandas as pd

entries = pd.read_parquet("hf://datasets/protplex/pdb-captions-2026/data/entries.parquet")
ligands = pd.read_parquet("hf://datasets/protplex/pdb-captions-2026/data/ligands.parquet")
bound = ligands[ligands.plip_class == "small_molecule"].merge(entries[["pdb_id", "title", "structural_tags"]], on="pdb_id")
-- DuckDB
SELECT e.pdb_id, e.title, p.volume_a3, p.druggability_score
FROM 'hf://datasets/protplex/pdb-captions-2026/data/entries.parquet' e
JOIN 'hf://datasets/protplex/pdb-captions-2026/data/pockets.parquet' p USING (pdb_id)
WHERE p.is_top_pocket AND e.assembly_composition = 'heteromer'
LIMIT 10;

What an entry contains

  • Texts. core_description, functional_annotations, biological_source, ligand_chemistry, experimental_details and literature describe the entry's biology and origin; summary_text joins them. structural_text describes the structure. These are the texts the Protplex search service returns for the entry, under the same field names, at full length.
  • Deterministic structural text. structural_tags and structural_text_deterministic are built from the measurements by fixed rules (vocabulary below). They are the inputs structural_text was written from.
  • Annotations (UniProtKB, GO, Pfam, EC, NCBI Taxonomy, CCD ligands) as identifier and name lists, and source metadata from the mmCIF file (title, method, resolution, dates, keywords, citation).
  • Measurements from seven structural tools: unmodelled residues and side-chain completeness, secondary structure, solvent-accessible surface, assembly composition and interfaces, predicted pockets, and protein-ligand interactions. Entry-level values are columns of entries; per-pocket, per-interface and per-ligand values are in their own tables.
  • Provenance for every value: tool, status, container image digest, the SHA-256 of the mmCIF file read, timestamps, and for each text the identifier of the instructions it was written under.

How the texts were made. The texts in summary_text, core_description, functional_annotations, biological_source, ligand_chemistry, experimental_details, literature, structural_text were written by a language model from the sources listed for each field (structural_text from structural_text_deterministic). They were not curated by hand and can contain statements that are not in those sources; Validation reports measured rates. structural_tags, structural_text_deterministic and every measurement are produced by fixed rules and tools.

Example (abridged), entry 9HEG:

pdb_id: 9HEG
title: Crystal structure of the oxidized respiratory complex I subunit NuoEF from Aquifex aeolicus, mutation V136M(NuoE)
core_description: The NuoEF subcomplex of respiratory complex I from Aquifex aeolicus with a V136M substitution; the subcomplex contains FMN and multiple iron–sulfur clusters that mediate electron transfer from NADH to the quinone pool. The structure highlights cofactors in the electron-transfer module.
structural_text: Heteromer with an obligate-like interface (Delta SASA ~4060 Angstrom^2) and an alpha-rich architecture. The compact assembly is helix-dominated with minimal disorder and a stable core. A cavernous, mixed-polarity high-confidence ligandable pocket is present, and a defined metal site with multiple cofactors is integrated into the electron-transfer module.
structural_tags: assembly:heteromer; interface:obligate-like; arch:alpha-rich; stability:compact; pocket:cavernous+mixed+high; metal_site:defined; ligand:CL; ligand:FMN; ligand:Fe-S_cluster; ligand:GOL; ligand:NA; ligand:SO4; ion:CL; ion:NA
structural_text_deterministic (clauses): heteromer with a obligate-like interface (largest Delta SASA ~4060 Angstrom^2); alpha-rich architecture; cavernous, mixed pocket with a ligandable site; defined metal site
assembly: heteromer A1B1, largest interface 4059.5 A^2 (both sides)
secondary structure: helix 0.46, strand 0.15
ligand_ids: FES, CL, SO4, SF4, FMN, GOL, NA

Selection and processing

Entries were selected by processing provenance, not by date or identifier: an entry is included when all seven structural tools have a result from the release's feature run at method version 2, and it passes the provenance gate (see Validation). Of the selected entries, 18,500 were first released between 2025-09-01 and 2026-09-30; 1,067 were released earlier (2020-09-02 onwards) but were not in Protplex's previous corpus, and were processed in the same run. They were processed from wwPDB files fetched on 2026-09-30 and SIFTS release 2026-09-27.

Step Tool Version Image digest Measures
parsers gemmi gemmi 0.7.5 in-process unmodelled residues, side-chain completeness, ligands and ions, structure preparation
dssp DSSP 4.4.0 sha256:4e941063537b secondary structure
freesasa FreeSASA assembly_sasa/2, gemmi 0.7.5 sha256:0e55f52f67d7 solvent-accessible surface of the asymmetric unit
assembly FreeSASA on biological assembly 1 (gemmi) assembly_sasa/2, gemmi 0.7.5 sha256:0e55f52f67d7 assembly composition and chain-pair interface areas
fpocket fpocket 4.2-seeded (seed 42) sha256:f778b055eb53 pocket detection and descriptors (fixed Monte Carlo seed)
p2rank P2Rank 2.5.1 sha256:ee00be29aae8 ligand-binding site prediction
plip PLIP 2.3.1 sha256:011b21c35375 protein-ligand interactions

The interface areas are measured on biological assembly 1 with FreeSASA (Lee-Richards, probe 1.4 A, heavy atoms, model 1, first altloc). Pockets come from fpocket (top three kept, fixed Monte Carlo seed so volumes are reproducible) and P2Rank (up to ten). PLIP classifies each ligand site by fixed rules and analyses interactions for small molecules, carbohydrates and metals; crystallisation and buffer additives are recorded but not analysed.

Structural vocabulary

Tags from symbolizer version 2. Each tag is a bin of one measured value; the thresholds are fixed.

Tag Values Rule
assembly monomer, homomer, heteromer Polymer entities of biological assembly 1 (of the asymmetric unit when no assembly result exists).
interface transient-like, moderate, large, obligate-like Bins of the largest interface's buried area (both sides): 200-800, 800-1,600, 1,600-3,000, >= 3,000 A^2. Area bins only: they do not classify an interface as transient or obligate.
arch alpha-rich, beta-rich, mixed_alpha/beta DSSP fractions: alpha-rich if helix >= 0.40 and helix - strand >= 0.15; beta-rich if strand >= 0.40 and strand - helix >= 0.15; otherwise mixed when helix + strand >= 0.10.
stability disulfide-rich, flexible, compact disulfide-rich: disulfides per protein residue >= 0.03, or >= 2 per chain with mean chain length <= 120; else flexible: internal unmodelled fraction >= 0.10 or an internal gap >= 50 residues; else compact: unmodelled fraction < 0.20.
model sidechains_truncated Fewer than half the protein residues have an atom beyond CB; pocket, PLIP, surface and interface values are then withheld from the tags.
pocket size+chemistry+ligandability The most druggable of fpocket's top three pockets (ties: larger volume). Size by volume: small < 300, medium < 700, large < 1,500, cavernous >= 1,500 A^3. Chemistry by apolar alpha-sphere proportion: polar < 0.40, mixed < 0.60, hydrophobic >= 0.60. Ligandability by druggability score: low < 0.40, medium < 0.60, high >= 0.60. Depth, burial and exposure are not measured.
na_binding present Protein-nucleic acid contacts measured by PLIP or the assembly interfaces, or nucleic-acid chains in the entry when neither measured.
metal_site defined PLIP finds a metal coordinated by the protein; without a PLIP result, a Zn, Mg, Mn, Fe, Ni, Cu or Ca ion is present.
ligand Ligand classes from the parser (e.g. metal:Ca2+) or CCD codes of bound ligands.
ion Ions present in the model.

The symbolizer also has fold, membrane and pore tags; no tool in this pipeline produces their inputs, so they never occur.

Columns

entries (one row per PDB entry)

Texts, annotations and entry-level structural measurements.

Column Type From Description
pdb_id string served PDB identifier (4 characters, upper case).
title string mmcif Title of the entry (_struct.title).
structural_text string served Paragraph describing the structure, written from the v2 structural measurements and tags of this entry (see structural_text_deterministic for those inputs). Where no paragraph was written, the deterministic text, as the service serves it (flag structural_text_deterministic).
summary_text string served The non-structural texts below joined in corpus order: the text the service embeds as the entry summary.
core_description string served Short description of the entry written from its title, UniProt names and functions, GO/Pfam/EC terms, ligand names, organism and citation title.
functional_annotations string served UniProt, GO, Pfam and EC annotations of the entry in prose form.
biological_source string served Source organism(s) in prose form.
ligand_chemistry string served Bound ligands, one line each: name, CCD code and a role.
experimental_details string served Method, resolution and crystallisation conditions in prose form.
literature string served Primary citation: title, journal and year, PubMed ID and DOI.
structural_tags string served Rule-based structural tags, "; "-separated (vocabulary in the card).
structural_text_deterministic string symbolizer The deterministic structural text built from the tool values by fixed rules: tags line, clause summary and key metrics. This is exactly what structural_text was written from.
structural_symbols string symbolizer The tags as a JSON object (assembly, interface, architecture, pocket size/chemistry/ligandability, ions, ligand classes, ...).
experimental_method_primary string served Primary experimental method.
resolution float64 served Resolution in angstrom (null for NMR and some other methods).
ph float64 served Crystallisation pH, when recorded.
temperature_k float64 served Crystallisation temperature in kelvin, when recorded.
deposition_date string mmcif Initial deposition date (YYYY-MM-DD).
initial_release_date string mmcif Date of the first public release (first revision in _pdbx_audit_revision_history).
latest_revision_date string mmcif Date of the latest revision present in the file the release was built from.
keywords string mmcif PDB keywords (_struct_keywords.pdbx_keywords).
pubmed_id string served PubMed ID of the primary citation.
organism_names list served Source organism names.
taxon_ids list served NCBI Taxonomy IDs of the source organisms.
tax_lineage list served Lineage strings of the source organisms (species > genus > family > order).
tax_lineage_ids list served NCBI Taxonomy IDs of every lineage node of the source organisms.
uniprot_ids list served UniProtKB accessions mapped to the entry's chains (SIFTS).
uniprot_names list served UniProtKB entry names, as "NAME (UniProt ACC)".
go_ids list served GO term IDs mapped to the entry's chains (SIFTS).
go_names list served GO term names, as "name (GO ID)".
pfam_ids list served Pfam family IDs mapped to the entry's chains (SIFTS, completed per chain).
pfam_names list served Pfam family names, as "name (Pfam ID)".
ec_ids list served EC numbers mapped to the entry's chains (SIFTS).
ec_names list served EC names (ENZYME).
ligand_ids list served CCD codes of the bound non-polymer components.
ligand_names list served CCD names, as "NAME (CODE)".
ligand_classes list served Normalised ligand classes present, assigned by fixed rules from the CCD (e.g. ion, metal, cofactor, drug_like, solvent, lipid, detergent).
ligand_class_map string served JSON object: CCD code -> its ligand classes.
entity_types list served Polymer entity types present (polypeptide(L), polydeoxyribonucleotide, ...).
seq_chain_ids list served Chain IDs of the polymer entities.
n_polymer_entities int64 served Number of polymer entities.
n_distinct_polymer_seqs int64 served Number of distinct polymer sequences.
n_protein_entities int64 served Number of protein entities.
missing_res_frac float64 parsers Fraction of polymer residues in the deposited sequence with no modelled coordinates (terminal and internal).
internal_missing_frac float64 parsers Fraction of polymer residues missing inside chains (gaps between modelled residues).
longest_internal_gap int64 parsers Longest run of unmodelled residues inside a chain.
protein_res_count int64 parsers Modelled protein residues.
na_res_count int64 parsers Modelled nucleic-acid residues.
sidechain_modeled_frac float64 parsers Fraction of protein residues with at least one atom beyond CB. Below 0.5 (backbone or CA traces) the pocket, PLIP, surface and interface values are kept out of the tags and texts; the raw values stay in these tables.
atom_completeness float64 parsers Modelled heavy atoms / heavy atoms expected for the modelled residues.
helix_frac float64 dssp Fraction of residues in helix (DSSP 4.4.0 classes H, G, I).
sheet_frac float64 dssp Fraction of residues in strand or bridge (DSSP classes E, B).
coil_frac float64 dssp Fraction of residues in neither.
helix_residues int64 dssp Residues in helix (H, G, I).
sheet_residues int64 dssp Residues in strand or bridge (E, B).
dssp_residues int64 dssp Residues DSSP assigned (model 1, first altloc).
disulfides int64 dssp Disulfide bonds annotated in _struct_conn (intra- and inter-chain, symmetry-generated ones excluded).
total_sasa_a2 float64 freesasa Solvent-accessible surface area of the asymmetric unit's polymer residues (angstrom^2; Lee-Richards, probe 1.4 A, heavy atoms, model 1).
buried_frac_au float64 freesasa Fraction of the chains' isolated surface buried between chains of the asymmetric unit.
assembly_id string assembly Biological assembly the interface values were measured on ("1").
assembly_source string assembly "assembly" when built from the file's assembly 1 operators; "asym_unit (...)" when the asymmetric unit was used instead, with the reason.
oligomeric_details string assembly Oligomeric description of assembly 1 as deposited (_pdbx_struct_assembly).
assembly_composition string assembly monomer, homomer or heteromer, from the polymer entities of assembly 1.
assembly_stoichiometry string assembly Stoichiometry of assembly 1 (A2, A2B2, ...).
assembly_chains int64 assembly Polymer chains in assembly 1.
n_interfaces int64 assembly Chain pairs in contact in assembly 1.
largest_interface_dsasa_a2 float64 assembly Buried surface of the largest chain-pair interface, both sides summed (angstrom^2). Per-side interface area, as RCSB and PISA report it, is half of this.
largest_interface_area_per_side_a2 float64 assembly Per-side area of the largest interface (angstrom^2).
total_buried_sasa_a2 float64 assembly Total surface buried on forming assembly 1 (angstrom^2).
largest_protein_na_dsasa_a2 float64 assembly Largest protein-nucleic acid interface, both sides (angstrom^2).
n_pockets_fpocket int64 fpocket Pockets fpocket found (all of them; the pockets table keeps the top 3).
n_druggable_pockets_fpocket int64 fpocket fpocket pockets with druggability score >= 0.5.
n_pockets_p2rank int64 p2rank P2Rank pockets kept (at most 10, in rank order).
plip_ligand_count int64 plip Distinct small-molecule or carbohydrate ligands (by name) with at least one PLIP interaction; additives, ions and nucleic acids are not counted.
plip_hbond_count int64 plip Hydrogen bonds, summed over the interacting small-molecule and carbohydrate sites.
plip_hydrophobic_count int64 plip Hydrophobic contacts, summed over the same sites.
plip_pistack_count int64 plip Pi-stacking interactions, summed over the same sites.
plip_metal_count int64 plip Metal coordination sites with at least one protein partner.
plip_has_nucleic_acid bool plip At least 3 protein residues in contact with nucleic acid.
status_parsers string release Outcome of the parsers step: done, skipped or error (the reason is in the provenance table). Null values of that tool's columns follow from it.
status_dssp string release Outcome of the dssp step: done, skipped or error (the reason is in the provenance table). Null values of that tool's columns follow from it.
status_freesasa string release Outcome of the freesasa step: done, skipped or error (the reason is in the provenance table). Null values of that tool's columns follow from it.
status_fpocket string release Outcome of the fpocket step: done, skipped or error (the reason is in the provenance table). Null values of that tool's columns follow from it.
status_p2rank string release Outcome of the p2rank step: done, skipped or error (the reason is in the provenance table). Null values of that tool's columns follow from it.
status_plip string release Outcome of the plip step: done, skipped or error (the reason is in the provenance table). Null values of that tool's columns follow from it.
status_assembly string release Outcome of the assembly step: done, skipped or error (the reason is in the provenance table). Null values of that tool's columns follow from it.
quality_flags list release Conditions a user may want to filter on (list in the card).

pockets (one row per predicted pocket)

Predicted pockets: the top 3 fpocket pockets and up to 10 P2Rank pockets per entry.

Column Type From Description
pdb_id string release PDB identifier.
method string release fpocket or p2rank.
rank int64 release Rank of the pocket in the tool's own order (1 = best).
score float64 release fpocket pocket score, or P2Rank pocket score.
probability float64 p2rank P2Rank ligandability probability.
druggability_score float64 fpocket fpocket druggability score (0-1).
volume_a3 float64 fpocket Pocket volume (angstrom^3; Monte Carlo with a fixed seed, reproducible).
n_alpha_spheres int64 fpocket Alpha spheres in the pocket.
apolar_alpha_sphere_proportion float64 fpocket Proportion of apolar alpha spheres: the value the pocket chemistry tag is binned from.
hydrophobicity_score float64 fpocket fpocket hydrophobicity score.
mean_local_hydrophobic_density float64 fpocket Mean local hydrophobic density.
polarity_score float64 fpocket fpocket polarity score.
charge_score float64 fpocket fpocket charge score.
proportion_polar_atoms_pct float64 fpocket Percentage of polar atoms.
total_sasa_a2 float64 fpocket Pocket solvent-accessible surface (angstrom^2).
polar_sasa_a2 float64 fpocket Polar part of it.
apolar_sasa_a2 float64 fpocket Apolar part of it.
mean_alpha_sphere_radius float64 fpocket Mean alpha sphere radius (angstrom).
mean_alpha_sphere_solvent_access float64 fpocket Mean alpha sphere solvent accessibility.
alpha_sphere_density float64 fpocket Alpha sphere density.
flexibility float64 fpocket Mean normalised B-factor of the pocket atoms.
sas_points int64 p2rank Solvent-accessible points in the P2Rank pocket.
surf_atoms int64 p2rank Surface atoms in the P2Rank pocket.
center_x float64 p2rank Pocket centre, x (angstrom, file coordinates).
center_y float64 p2rank Pocket centre, y.
center_z float64 p2rank Pocket centre, z.
n_residues int64 release Residues lining the pocket.
residues list release Lining residues as CHAIN_NUMBER[INSCODE][:RESNAME] (fpocket) or CHAIN_NUMBER (P2Rank), with author chain IDs and residue numbers.
is_top_pocket bool symbolizer The fpocket pocket the pocket tag describes (the most druggable of the top three; ties: larger volume).

interfaces (one row per chain pair in contact)

Chain-pair interfaces of biological assembly 1.

Column Type From Description
pdb_id string release PDB identifier.
assembly_id string assembly Assembly ("1").
chain_1 string assembly First chain (assembly chain name).
chain_2 string assembly Second chain (assembly chain name).
source_chain_1 string assembly Author chain ID the first chain was generated from.
source_chain_2 string assembly Author chain ID the second chain was generated from.
kind string assembly protein-protein, protein-na, na-na or other.
dsasa_a2 float64 assembly Buried surface, both sides summed (angstrom^2).
area_per_side_a2 float64 assembly Per-side interface area (angstrom^2).
atoms_in_contact int64 assembly Atoms in contact across the interface.

ligands (one row per PLIP binding site)

Bound ligands and their measured interactions (PLIP 2.3.1).

Column Type From Description
pdb_id string release PDB identifier.
site_id string plip PLIP binding-site ID (CCD:CHAIN:NUMBER).
ccd_id string plip CCD code of the ligand.
ccd_name string served CCD name of the ligand.
chain string plip Author chain ID of the ligand.
residue_number string plip Author residue number of the ligand.
plip_class string plip Site class by fixed rules: small_molecule, carbohydrate, glycosylation, metal (ions), nucleic_acid, or additive (a fixed list of crystallisation and buffer components and their ions).
ligand_classes list served The entry-level ligand classes of this CCD code.
ligtype string plip PLIP ligand type.
covalent bool plip Covalently bound to the polymer.
composite bool plip Composite ligand (several linked components).
n_heavy_atoms int64 plip Heavy atoms of the ligand.
has_interactions bool plip Whether PLIP found any interaction.
interacting_chains list plip Chains the ligand interacts with.
n_hydrophobic int64 plip Hydrophobic interactions (null when the site was not analysed).
n_hbond int64 plip Hbond interactions (null when the site was not analysed).
n_water_bridge int64 plip Water bridge interactions (null when the site was not analysed).
n_salt_bridge int64 plip Salt bridge interactions (null when the site was not analysed).
n_pi_stack int64 plip Pi stack interactions (null when the site was not analysed).
n_pi_cation int64 plip Pi cation interactions (null when the site was not analysed).
n_halogen int64 plip Halogen interactions (null when the site was not analysed).
n_metal int64 plip Metal interactions (null when the site was not analysed).
contact_residues list plip Interacting residues as CHAIN:RESNAMENUMBER.
metal_sites string plip JSON list of metal coordination sites (metal, geometry, coordination number, partners).
smiles string plip SMILES of the ligand as PLIP read it.
inchikey string plip InChIKey of the ligand as PLIP read it.

provenance (one row per entry and processing step)

How and from what each value was computed.

Column Type From Description
pdb_id string release PDB identifier.
step string release A structural tool (parsers, dssp, freesasa, fpocket, p2rank, plip, assembly) or a text stage (text:unify for the non-structural texts, text:structural_refine for structural_text).
status string release done, skipped or error.
reason string release Why a step was skipped or failed.
method_version int64 release Method version of the tool step (2 for every released row).
image string release Container image the tool ran in.
image_digest string release Content digest of that image.
input_sha256 string release SHA-256 of the mmCIF file the tool read.
finished_at string release When the step finished (UTC, ISO 8601).
seconds float64 release Run time of the step.
prompt_version string prose-store Identifier of the instructions a text stage was written under.
input_digest_match bool prose-store For text:structural_refine: whether the paragraph's recorded input digest equals the digest of this entry's structural inputs (true for every released row).

"From": served = as the Protplex service returns it for this entry; mmcif = the entry's PDBx/mmCIF file; a tool name = computed by that tool; symbolizer = fixed-rule vocabulary built from the tool values; prose-store = the record of the text stage; release = derived by the release export.

Quality flags

quality_flags lists conditions a user may want to filter on.

Flag Meaning Entries
dssp_skipped dssp did not run: e.g. "1198 MB mmCIF: above the 250 MB limit" (2) 2
freesasa_skipped freesasa did not run: e.g. "1198 MB mmCIF: above the 250 MB limit" (2) 2
fpocket_skipped fpocket did not run: e.g. "no PDB-format input: 86 chains > 62" (339); e.g. "143680 heavy atoms: above the fpocket limit of 100000" (270); e.g. "1198 MB mmCIF: above the 250 MB limit" (2) 611
p2rank_skipped p2rank did not run: e.g. "1198 MB mmCIF: above the 250 MB limit" (2) 2
plip_skipped plip did not run: e.g. "1198 MB mmCIF: above the 250 MB limit" (2) 2
assembly_skipped assembly did not run: e.g. "1198 MB mmCIF: above the 250 MB limit" (2) 2
fpocket_error fpocket failed: e.g. "fpocket: timed out after 300 s" (11) 11
plip_error plip failed: "plip error: ValueError: invalid literal for int() with base 10: 'NA'" (1); "plip partial: batch 1: plip rc=124: timeout" (1) 2
structural_text_deterministic No paragraph was written for this entry; structural_text (as served) is the deterministic text. 4
no_structural_text structural_text is null; structural_text_deterministic is still present. 0
sidechains_truncated Fewer than half the protein residues have an atom beyond CB: the pocket, PLIP, surface and interface values are kept out of the tags and texts (the tables keep them). 116
interfaces_truncated The interface list of assembly 1 was capped (very large assemblies). 145
method_not_parsed experimental_method_primary holds a raw mmCIF token instead of a method name. 18
doi_missing The literature text carries the placeholder "DOI:?". 5,096
no_uniprot_mapping No UniProtKB accession is mapped to the entry. 1,979

Validation

  • Processing provenance (pass): 19,567 entries released, all with complete v2 provenance.
  • Assembly claims against RCSB (measured): structural_text oligomeric label wrong in 0/299 sampled entries after adjudication.
  • Pocket claims (measured): tagged pocket touches a bound ligand in 223/300 sampled ligand-bound entries.
  • Unmeasured terms in structural_text (measured): 16,071 of 19,567 structural paragraphs use a term no measurement backs.
  • Ligand roles (measured): 4,577 of 30,721 ligand roles name a PLIP-classed additive.
  • Dataset texts compared with the service (measured): 19,439 of 19,567 entries have fuller texts here than in the service.
  • Record read against source facts (measured): 100 records read; highest unsupported-claim rate 97%.

Details, sample sizes and confidence intervals are in VALIDATION.md; machine-readable results are in validation/.

Limitations

  • Coverage. 19,567 entries: the entries the v2 pipeline has processed so far, not the whole PDB. 18,500 were first released between 2025-09-01 and 2026-09-30; 1,067 were released earlier (2020-09-02 onwards) but were not in Protplex's previous corpus, and were processed in the same run. The method mix follows recent depositions (X-ray crystallography 10,708, Electron microscopy 8,555, NMR spectroscopy 286). The rest of the PDB will follow once it is reprocessed.
  • Texts. summary_text, core_description, functional_annotations, biological_source, ligand_chemistry, experimental_details, literature, structural_text are written by a language model from the entry's source records and measurements, not curated by hand, and can contain statements those sources do not support (rates under Validation).
  • Structural paragraphs. 16071/19567 (82.1%, 95% CI 81.6-82.7%) of structural_text paragraphs use at least one term no measurement in the release backs (fold names, pocket exposure or depth, grooves, ...); structural_text_deterministic uses only the tag vocabulary.
  • Interface bins are area bins. "Obligate-like" and "transient-like" name ranges of buried area; they are not a classification of interface biology.
  • Pockets. The pocket tag describes fpocket's most druggable top-three pocket, which need not be the ligand site (in entries with a bound small molecule, it touches a ligand in 223/300 (74.3%, 95% CI 69.1-79.0%)); pocket depth, burial and exposure are not measured.
  • Ligand roles. 95% of ligand_chemistry lines end with a role (structural cofactor, regulatory ion, catalytic cofactor, inhibitor, substrate analog, peptide or nucleic acid ligand, and a few others); no source record supplies roles, and 4577/30721 (14.9%, 95% CI 14.5-15.3%) of them are given to components PLIP classes as crystallisation or buffer additives.
  • Annotation snapshots. Chain mappings come from SIFTS 2026-09-27; names and function text from UniProtKB/Swiss-Prot (2025-06-18), GO (2025-07-22), Pfam (2025-07-02) and ENZYME (18-Jun-2025), which predate many of the structures: an identifier added since has no name. 11,231 entries (57%) have a GO term; the SIFTS release used carries far fewer keyword-derived GO terms than earlier releases.
  • Missing values. fpocket did not run on 611 entries (more than 62 chains or 100,000 heavy atoms) and failed on 11; the coordinate tools did not run on 2 entries with mmCIF files over 250 MB; 116 models lack most side chains, and their pocket, interaction and surface values are kept out of the tags. Null values follow the status_<tool> columns.
  • Literature is the primary citation only (title, journal, year, PubMed ID, DOI); no abstracts or full text. 5,096 entries carry the placeholder "DOI:?".
  • 18 entries have an unparsed method token in experimental_method_primary (flag method_not_parsed).
  • Not a benchmark. There is no train/test split. For learning tasks, split by UniProt accession or sequence cluster: many entries are structures of the same protein.

Sources and licences

The dataset is released under the Creative Commons Attribution 4.0 licence (CC BY 4.0). It is derived from the sources below; please cite them, and respect their own terms, when you use the corresponding fields.

Source Supplies Version used Licence Cite
wwPDB archive coordinates, entry metadata, titles, citations, dates 2026-09-30 CC0 1.0 wwPDB consortium (2019) Protein Data Bank: the single global archive for 3D macromolecular structure data. Nucleic Acids Res 47:D520-D528.
Chemical Component Dictionary ligand names and chemistry 2026-09-26 CC0 1.0 Westbrook JD et al. (2015) The chemical component dictionary. Bioinformatics 31:1274-1278.
SIFTS chain-level mappings to UniProtKB, GO, Pfam, EC and NCBI Taxonomy 2026-09-27 CC0 1.0 (PDBe public data) Dana JM et al. (2019) SIFTS: updated Structure Integration with Function, Taxonomy and Sequences resource. Nucleic Acids Res 47:D482-D489.
UniProtKB/Swiss-Prot protein names and function text 2025-06-18 CC BY 4.0 The UniProt Consortium (2023) UniProt: the Universal Protein Knowledgebase in 2023. Nucleic Acids Res 51:D523-D531.
Gene Ontology GO term names 2025-07-22 CC BY 4.0 Gene Ontology Consortium (2023) The Gene Ontology knowledgebase in 2023. Genetics 224:iyad031.
Pfam Pfam family names 2025-07-02 CC0 1.0 Mistry J et al. (2021) Pfam: The protein families database in 2021. Nucleic Acids Res 49:D412-D419.
ENZYME EC names 18-Jun-2025 CC BY 4.0 Bairoch A (2000) The ENZYME database in 2000. Nucleic Acids Res 28:304-305.
NCBI Taxonomy organism names and lineages 2025-09-10 public domain (NCBI); attribution requested Schoch CL et al. (2020) NCBI Taxonomy: a comprehensive update on curation, resources and tools. Database 2020:baaa062.

Tools (please cite when you use their values):

  • gemmi: Wojdyr M (2022) GEMMI: A library for structural biology. J Open Source Softw 7:4200.
  • DSSP: Kabsch W, Sander C (1983) Dictionary of protein secondary structure. Biopolymers 22:2577-2637.
  • FreeSASA: Mitternacht S (2016) FreeSASA: An open source C library for solvent accessible surface area calculations. F1000Research 5:189.
  • FreeSASA on biological assembly 1 (gemmi): Mitternacht S (2016) F1000Research 5:189; Wojdyr M (2022) J Open Source Softw 7:4200.
  • fpocket: Le Guilloux V, Schmidtke P, Tuffery P (2009) Fpocket: an open source platform for ligand pocket detection. BMC Bioinformatics 10:168.
  • P2Rank: Krivak R, Hoksza D (2018) P2Rank: machine learning based tool for rapid and accurate prediction of ligand binding sites from protein structure. J Cheminform 10:39.
  • PLIP: Adasme MF et al. (2021) PLIP 2021: expanding the scope of the protein-ligand interaction profiler to DNA and RNA. Nucleic Acids Res 49:W530-W534.

Versions

Version Date Changes
0.1.0 2026-11-12 First release: entries processed by the v2 structural pipeline.

Each release is a tagged revision of this repository (revision="v<version>"). PROVENANCE.json records the code version, container image digests, source versions and file checksums of this release.

Citation

If you use this dataset, please cite it, and also cite the sources and tools whose fields you use (listed under Sources and licences).

@misc{protplex_pdb_captions_2026,
  author    = {Pariset, Matteo},
  title     = {Protplex PDB Captions (2026 Slice)},
  year      = {2026},
  version   = {0.1.0},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/protplex/pdb-captions-2026}
}

Matteo Pariset (2026). Protplex PDB Captions (2026 Slice) (version 0.1.0) [Data set]. Hugging Face. https://huggingface.co/datasets/protplex/pdb-captions-2026