|
Download VALIDATION.md from protplex/pdb-captions-2026: direct link, hf CLI and curl.
- Browser
- Download file 15.3 kB
-
https://huggingface.co/datasets/protplex/pdb-captions-2026/resolve/main/VALIDATION.md
- Command line
-
hf download hf://datasets/protplex/pdb-captions-2026/VALIDATION.md
-
curl -L -o VALIDATION.md https://huggingface.co/datasets/protplex/pdb-captions-2026/resolve/main/VALIDATION.md
15.3 kB
| # Validation and provenance: Protplex PDB Captions (2026 Slice), version 0.1.0 | |
| What was checked for this release, how, and with what result. Rates are given with 95% Wilson confidence intervals. Machine-readable results are in `validation/<check>.json`. | |
| ## Processing provenance | |
| Status: **pass**. 19,567 entries released, all with complete v2 provenance. | |
| - 19,567 entries have method-version-2 results for all 7 tools from the release's feature run; 19,567 pass the provenance gate. | |
| - Gate rules, applied to every entry: each tool's result comes from the pinned container image; every tool read the mmCIF file that is on disk (same SHA-256), and the corpus parsed the same file; the structural text uses no 2025 (pre-v2) result; the served tags equal the tags the feature database gives; each structural paragraph is the stored one, and its recorded input digest equals the digest of the entry's current structural inputs. | |
| - 19,563 of 19,567 structural paragraphs were written from exactly the measurements in this release. | |
| - The feature run's code digests match commit fa427cb and the current code. | |
| - Tool outcomes: parsers done 19,567; dssp done 19,565, skipped 2; freesasa done 19,565, skipped 2; fpocket done 18,945, error 11, skipped 611; p2rank done 19,565, skipped 2; plip done 19,563, error 2, skipped 2; assembly done 19,565, skipped 2. | |
| ## Assembly claims against RCSB | |
| Status: **measured**. structural_text oligomeric label wrong in 0/299 sampled entries after adjudication. | |
| Random sample of 300 released entries with a structural paragraph (seed 20261111); reference: RCSB assembly 1, fetched 2026-11-11T18:59:17Z (299 entries with an assembly 1). Each flagged case was inspected; where the flag is a misreading by the claim extractor or a gap in the reference, the count after adjudication excludes it (reasons in validation/rcsb_assembly.json). | |
| - `structural_text`: oligomeric label wrong 1/299 (0.3%, 95% CI 0.1-1.9%), 0 after adjudication; subunit count wrong 1/174 (0.6%, 95% CI 0.1-3.2%), 0 after adjudication; nucleic-acid claim with no nucleic acid in assembly 1 3/35 (8.6%, 95% CI 3.0-22.4%), 1 after adjudication; metal claim with no metal in the entry 4/96 (4.2%, 95% CI 1.6-10.2%), 0 after adjudication. | |
| Quoted interface areas are a median 2.014x RCSB's largest interface area (174/181 between 1.6x and 2.5x): the text quotes the area buried on both sides, RCSB reports one side. | |
| "Obligate-like" (an area bin, >= 3,000 A^2 both sides) where RCSB's largest interface is below 800 A^2: 3/133 (2.3%, 95% CI 0.8-6.4%). | |
| - `core_description`: oligomeric label wrong 3/21 (14.3%, 95% CI 5.0-34.6%), 1 after adjudication; subunit count wrong 11/31 (35.5%, 95% CI 21.1-53.0%), 1 after adjudication. | |
| - Measured values: `assembly_composition` wrong 0/299 (0.0%, 95% CI 0.0-1.3%); `assembly_chains` wrong 0/299 (0.0%, 95% CI 0.0-1.3%); `largest_interface_area_per_side_a2` is a median 1.007x RCSB's largest interface (196/205 within 10%). | |
| - Claims are extracted by fixed patterns (oligomer words, "nucleic"/"DNA"/"RNA", "metal", "Delta SASA ~N"); the metal check is a necessary condition only (the entry contains a metal). | |
| ## Pocket claims | |
| Status: **measured**. tagged pocket touches a bound ligand in 223/300 sampled ligand-bound entries. | |
| Text against tags, all 18,729 entries with a pocket tag: | |
| - pocket size stated differently from the tag: 113/18285 (0.6%, 95% CI 0.5-0.7%); not stated: 444/18729 (2.4%, 95% CI 2.2-2.6%). | |
| - pocket chemistry stated differently: 24/18065 (0.1%, 95% CI 0.1-0.2%); not stated: 664/18729 (3.5%, 95% CI 3.3-3.8%). | |
| - a low-ligandability pocket called ligandable without a qualifier: 436/8029 (5.4%, 95% CI 5.0-5.9%); a high-ligandability pocket described as weakly ligandable: 0/8755 (0.0%, 95% CI 0.0-0.0%). | |
| - a pocket mentioned although the entry has no pocket tag: 557/838 (66.5%, 95% CI 63.2-69.6%). | |
| Chemistry tag against lining residues (18,729 tagged pockets): median share of hydrophobic residues (A, V, L, I, M, F, W, P, C) by chemistry bin: hydrophobic: 0.603 (n = 6,382); mixed: 0.429 (n = 5,775); polar: 0.318 (n = 6,572); Spearman correlation between the apolar alpha-sphere proportion and that share: 0.694. | |
| Pocket against bound ligands, from the deposited coordinates (300 of 300 sampled entries with a small-molecule or carbohydrate ligand of at least 6 heavy atoms; contact = polymer residue within 4.0 A): | |
| - the tagged pocket shares a residue with a ligand site: 223/300 (74.3%, 95% CI 69.1-79.0%); by ligandability bin: high 153/181 (84.5%, 95% CI 78.5-89.1%); low 52/93 (55.9%, 95% CI 45.8-65.6%); medium 18/26 (69.2%, 95% CI 50.0-83.5%). | |
| - any of fpocket's top three pockets: 268/300 (89.3%, 95% CI 85.3-92.3%); P2Rank's first pocket: 204/296 (68.9%, 95% CI 63.4-73.9%). | |
| - The pocket tag describes fpocket's most druggable top-three pocket, not necessarily the ligand site; pocket depth, burial and exposure are not measured, so words such as "surface", "enclosed" or "buried" in structural_text are not backed by a measurement (see Unmeasured terms). | |
| ## Unmeasured terms in structural_text | |
| Status: **measured**. 16,071 of 19,567 structural paragraphs use a term no measurement backs. | |
| All 19,567 released structural paragraphs. "Unbacked" = the paragraph uses the term and the entry has no tag that measures it. | |
| - fold name: used in 7580/19567 (38.7%, 95% CI 38.1-39.4%); unbacked in 7580/19567 (38.7%, 95% CI 38.1-39.4%) (no fold assignment is computed). | |
| - pocket exposure or depth: used in 9652/19567 (49.3%, 95% CI 48.6-50.0%); unbacked in 9652/19567 (49.3%, 95% CI 48.6-50.0%) (pocket depth and burial are not measured). | |
| - groove: used in 2658/19567 (13.6%, 95% CI 13.1-14.1%); unbacked in 2658/19567 (13.6%, 95% CI 13.1-14.1%) (not measured). | |
| - pore or channel: used in 304/19567 (1.6%, 95% CI 1.4-1.7%); unbacked in 304/19567 (1.6%, 95% CI 1.4-1.7%) (no pore tool runs). | |
| - membrane topology: used in 589/19567 (3.0%, 95% CI 2.8-3.3%); unbacked in 589/19567 (3.0%, 95% CI 2.8-3.3%) (no membrane tool runs). | |
| - catalytic or active site: used in 1683/19567 (8.6%, 95% CI 8.2-9.0%); unbacked in 1683/19567 (8.6%, 95% CI 8.2-9.0%) (not measured). | |
| - allosteric: used in 27/19567 (0.1%, 95% CI 0.1-0.2%); unbacked in 27/19567 (0.1%, 95% CI 0.1-0.2%) (not measured). | |
| - conformational state: used in 17/19567 (0.1%, 95% CI 0.1-0.1%); unbacked in 17/19567 (0.1%, 95% CI 0.1-0.1%) (not measured). | |
| - stable or rigid core: used in 651/19567 (3.3%, 95% CI 3.1-3.6%); unbacked in 651/19567 (3.3%, 95% CI 3.1-3.6%) (not measured). | |
| - compact: used in 17210/19567 (87.9%, 95% CI 87.5-88.4%); unbacked in 3057/19567 (15.6%, 95% CI 15.1-16.1%) (tag stability:compact). | |
| - nucleic-acid contacts: used in 2672/19567 (13.7%, 95% CI 13.2-14.1%); unbacked in 558/19567 (2.9%, 95% CI 2.6-3.1%) (tag na_binding:present). | |
| - metal site: used in 6970/19567 (35.6%, 95% CI 34.9-36.3%); unbacked in 899/19567 (4.6%, 95% CI 4.3-4.9%) (tag metal_site:defined). | |
| - obligate-like interface: used in 8455/19567 (43.2%, 95% CI 42.5-43.9%); unbacked in 134/19567 (0.7%, 95% CI 0.6-0.8%) (tag interface:obligate-like (an area bin)). | |
| - Paragraphs with at least one unbacked term: 16071/19567 (82.1%, 95% CI 81.6-82.7%). The deterministic text (structural_text_deterministic) uses only tag vocabulary. | |
| ## Ligand roles | |
| Status: **measured**. 4,577 of 30,721 ligand roles name a PLIP-classed additive. | |
| - 30721/32471 (94.6%, 95% CI 94.4-94.8%) of the 32,471 ligand lines end with a role. After normalising the wording (prefixes, trailing full stops), the roles are: structural cofactor 8,375, regulatory ion 6,462, catalytic cofactor 5,579, inhibitor 4,965, substrate analog 3,655, other 1,221, peptide ligand 299, nucleic acid ligand 165; 1221/30721 (4.0%, 95% CI 3.8-4.2%) fall outside the seven labels. None of the source records supplies a role field. | |
| - 4577/30721 (14.9%, 95% CI 14.5-15.3%) of the roles are given to a component that PLIP's fixed rules class as a crystallisation or buffer additive. Most frequent: SO4 as regulatory ion (636), GOL as structural cofactor (456), PO4 as regulatory ion (344), SO4 as structural cofactor (279), EDO as structural cofactor (257), PEG as structural cofactor (207), ACT as regulatory ion (178), Y01 as structural cofactor (135). | |
| - For ligand identity, use `ligand_ids`, `ligand_names` and the ligands table (PLIP site class, measured interactions); the class there is rule-based, not a biological role. | |
| ## Dataset texts compared with the service | |
| Status: **measured**. 19,439 of 19,567 entries have fuller texts here than in the service. | |
| The service builds its index with three economies the dataset does not apply: a bucket text shorter than 6 words is dropped (except core_description and structural_features), a text identical to another entry's is kept only on the entry with the smallest PDB ID, and the stored payload caps each bucket text at 800 characters and the summary and structural texts at 2,000. | |
| - Released entries whose served bucket texts differ from the dataset's: 19439/19567 (99.4%, 95% CI 99.2-99.5%); summary_text differs: 19436/19567 (99.3%, 95% CI 99.2-99.4%); structural_text truncated by the service: 0/19567 (0.0%, 95% CI 0.0-0.0%). | |
| - `core_description`: present in 19,567 entries; the service drops it as too short in 0, as shared with another entry in 371, and truncates it in 0. | |
| - `functional_annotations`: present in 17,162 entries; the service drops it as too short in 202, as shared with another entry in 3,567, and truncates it in 366. | |
| - `biological_source`: present in 19,560 entries; the service drops it as too short in 14,774, as shared with another entry in 2,432, and truncates it in 0. | |
| - `ligand_chemistry`: present in 15,028 entries; the service drops it as too short in 5,377, as shared with another entry in 2,351, and truncates it in 2. | |
| - `experimental_details`: present in 19,567 entries; the service drops it as too short in 15,215, as shared with another entry in 890, and truncates it in 0. | |
| - `literature`: present in 19,567 entries; the service drops it as too short in 0, as shared with another entry in 9,618, and truncates it in 0. | |
| ## Record read against source facts | |
| Status: **measured**. 100 records read; highest unsupported-claim rate 97%. | |
| 100 randomly sampled entries (seed 20261113); each text judged against the entry's source facts (mmCIF fields, UniProt function text, GO/Pfam/EC names, organisms, ligands, and for structural_text the deterministic text and measurements). Verdicts: supported = every statement follows from the source facts; embellished = adds interpretive or generic wording not in the sources, with no specific checkable claim; unsupported = makes at least one specific factual claim the sources do not contain; contradicted = states something the sources contradict. Reviewers judged against the facts only; for each claim not in the facts they also recorded whether they believe it (textbook knowledge, plausible inference, doubtful, false). The reviewers were language-model agents working to a written rubric, one per entry. In a second automated pass, all 15 contradicted verdicts and 10 random others were re-read against their packets; 1 verdict was changed (listed below). There was no human annotation and no double annotation, so these rates are indicative. | |
| - `core_description`: unsupported or contradicted 82/100 (82.0%, 95% CI 73.3-88.3%), of which only textbook knowledge in 20; a claim the reviewer judged doubtful or false, or a contradiction, 8/100 (8.0%, 95% CI 4.1-15.0%) (contradicted 1/100 (1.0%, 95% CI 0.2-5.5%)); embellished 0/100 (0.0%, 95% CI 0.0-3.7%). | |
| - `functional_annotations`: unsupported or contradicted 34/89 (38.2%, 95% CI 28.8-48.6%), of which only textbook knowledge in 23; a claim the reviewer judged doubtful or false, or a contradiction, 3/89 (3.4%, 95% CI 1.1-9.4%) (contradicted 1/89 (1.1%, 95% CI 0.2-6.1%)); embellished 2/89 (2.2%, 95% CI 0.6-7.8%). | |
| - `biological_source`: unsupported or contradicted 30/100 (30.0%, 95% CI 21.9-39.6%), of which only textbook knowledge in 26; a claim the reviewer judged doubtful or false, or a contradiction, 3/100 (3.0%, 95% CI 1.0-8.5%) (contradicted 0/100 (0.0%, 95% CI 0.0-3.7%)); embellished 0/100 (0.0%, 95% CI 0.0-3.7%). | |
| - `ligand_chemistry`: unsupported or contradicted 9/77 (11.7%, 95% CI 6.3-20.8%), of which only textbook knowledge in 1; a claim the reviewer judged doubtful or false, or a contradiction, 2/77 (2.6%, 95% CI 0.7-9.0%) (contradicted 1/77 (1.3%, 95% CI 0.2-7.0%)); embellished 0/77 (0.0%, 95% CI 0.0-4.8%). | |
| - `experimental_details`: unsupported or contradicted 0/100 (0.0%, 95% CI 0.0-3.7%), of which only textbook knowledge in 0; a claim the reviewer judged doubtful or false, or a contradiction, 0/100 (0.0%, 95% CI 0.0-3.7%) (contradicted 0/100 (0.0%, 95% CI 0.0-3.7%)); embellished 0/100 (0.0%, 95% CI 0.0-3.7%). | |
| - `literature`: unsupported or contradicted 0/100 (0.0%, 95% CI 0.0-3.7%), of which only textbook knowledge in 0; a claim the reviewer judged doubtful or false, or a contradiction, 0/100 (0.0%, 95% CI 0.0-3.7%) (contradicted 0/100 (0.0%, 95% CI 0.0-3.7%)); embellished 0/100 (0.0%, 95% CI 0.0-3.7%). | |
| - `structural_text`: unsupported or contradicted 97/100 (97.0%, 95% CI 91.5-99.0%), of which only textbook knowledge in 0; a claim the reviewer judged doubtful or false, or a contradiction, 30/100 (30.0%, 95% CI 21.9-39.6%) (contradicted 11/100 (11.0%, 95% CI 6.2-18.6%)); embellished 3/100 (3.0%, 95% CI 1.0-8.5%). | |
| - Second pass: 1 verdicts changed, by rule: second pass: the reviewer compared ligandability with P2Rank instead of the fpocket tag (1). | |
| ## Earlier results on the same pipeline and entries | |
| Reported by earlier validation work; not re-run for this release. | |
| | Check | Result | Source | | |
| |---|---|---| | |
| | Oligomeric label in v2-based structural paragraphs vs RCSB assembly 1 | 0/300 wrong | Protplex publication-feasibility audit, 2026-11-09 (random sample of the same entries) | | |
| | Subunit count in those paragraphs vs RCSB | 0/148 wrong | same audit | | |
| | Nucleic-acid contact claims vs RCSB composition | 0/36 false | same audit | | |
| | Quoted interface area vs RCSB largest interface | 2.0x (both sides vs one) in 189/198 | same audit | | |
| | Largest interface area per side vs RCSB | within 5% on 197/198 entries | v2 pipeline validation, 2026-09/10 | | |
| | Unmodelled residues vs the mmCIF's own records | equal on 12,750/12,750 chains | v2 pipeline validation, 2026-09/10 | | |
| | Side-chain completeness vs an independent gemmi computation | 1,000/1,000 entries equal | v2 pipeline validation, 2026-09 | | |
| | fpocket volumes on rerun with the pinned image and seed | identical between two runs on 22 hard-case entries | v2 pipeline rehearsals, 2026-09 (final code) | | |
| ## Not validated | |
| - Secondary-structure fractions and solvent-accessible areas were not compared with an external reference in this release. | |
| - PLIP interactions were not compared with an external reference. | |
| - fpocket and P2Rank pockets are predictions; only their overlap with bound ligands is measured here. | |
| - Ligand roles in ligand_chemistry have no source field to compare with. | |
| - The non-structural texts were read against their sources on a sample only (Record read). | |
| This release is not described as scientifically validated as a whole: the checks above measure specific claims and values, on the samples stated. | |