PDB: Deposited Structures as Functional-Insight Evidence

Bottom line: a deposited structure that captures a bound cofactor, ligand or
partner is direct experimental evidence for molecular function and complex
membership. We inventoried the PDB cross-references of every gene in the
pipeline (949 of 2,529 genes, 13,415 entries), ranked candidates by structure
richness times annotation sparsity, and measured how often structure papers
enter the GO evidence trail: only 15% of 737 structure-paper and gene pairs are
cited in GOA. We then tested whether structures fill gaps that ordinary papers
would not. They reliably give under-curated proteins their first
experimental-grade evidence (about 12 annotations in round 2, and four NEW
annotations across mcrA, secA and HSPB3), but new informative function is rare:
two structure-unique annotations, both on merA, and one clear gain in six when
scoring the papers' headline hypotheses. A first batch of six eukaryotic reviews
and four frontier reviews (merA, mcrA, secA, mxaI) used structure evidence; the
ranked worklist in PDB/GAP_WORKLIST.md is the next step.

We did this because structures are an under-used evidence source for GO, and we
wanted to know where structure-informed review pays off before spending effort
on it. The main limits turned out to be subunit mismatch (the paper is about a
partner) and GO expressivity (no term for the function the paper shows).

Goal

Many genes in this review pipeline have experimentally solved 3D structures deposited in
the Protein Data Bank. A structure is not just a picture: when it
captures a bound ligand, cofactor, catalytic metal, or partner macromolecule, it is
direct, experimental evidence for molecular function, binding, and complex membership — the
very things GO annotation review hinges on.

This project (1) inventories what is deposited across the whole pipeline, and (2) prioritizes
the structures most likely to change or strengthen a gene's functional annotation, so that
structure-informed review effort goes where it pays off.

What "functional insight" means here

A deposited structure is most valuable for review when what is bound in it tells us
something the current annotations do not. The prioritization therefore combines two axes:

1. Structure richness — what the structure actually contains:
2. bound cofactor / catalytic metal (FAD, NAD(P), PQQ, F430, Fe–S clusters, heme, Zn, …)
   → strongest single clue to catalytic mechanism / molecular function
3. bound substrate / product / inhibitor ligand → active-site and specificity evidence
4. complex (≥2 protein entities, or bound nucleic acid) → interaction & CC evidence

5. high sequence coverage (full-length vs a short peptide fragment in a partner's structure)

6. Annotation sparsity — a structure adds the most where we currently know the least.
   Genes whose GO annotations are electronic-only (IEA/IBA) with no experimental
   molecular-function term are upweighted: here a ligand-bound structure (plus literature)
   can ground a specific, experiment-grade MF annotation that today rests only on inference.

The two are multiplied: richness × sparsity. An apo structure of an already
well-annotated protein scores low; a cofactor-bound, full-length structure of an IEA-only
enzyme scores high.

What is deposited (pipeline-wide inventory)

Computed offline from the cached UniProt records (*-uniprot.txt, DR PDB cross-references):

Top organisms by genes-with-structure: human (588), yeast (65), ARATH (53), BPT4 (32),
mouse (30), ECOLI (23), PSEPK (19), SCHPO (17), worm (17), rat (15).

Three prioritization cuts

enrich_rcsb.py enriches the union of three candidate cuts (278 genes; each tagged in the
candidate_reason column), so the priority list is no longer prokaryote-dominated:

Full inventory: PDB/data/pdb_inventory.tsv (per entry) and
pdb_gene_summary.tsv (per gene).

Prioritized candidates (first pass)

The 114 genes with a structure but no experimental MF GO were enriched with RCSB metadata
(bound ligands, entity counts, titles). Of the 105 that resolved:

• 42 have a structure with a bound cofactor / catalytic metal
• 69 have a structure with a meaningful (non-buffer) ligand
• 40 are captured in a complex; 6 with bound nucleic acid
• 23 are apo with no partner (lower structural-insight yield)

These are IEA-only enzymes/proteins whose function is structurally (and often
biochemically) characterized but whose GO annotations have not been promoted beyond
electronic inference — the sweet spot for structure-grounded review.

Full ranked list: PDB/data/pdb_gene_enriched.tsv (now includes the
RCSB per-entry structure-paper PMIDs and an is_eukaryote flag).

Eukaryotic candidates (so they aren't drowned out)

Broadening beyond the strict "no experimental MF" cut to euk + contested surfaces a
much richer eukaryotic slice (815 eukaryotic genes have a structure). Top eukaryotic
candidates by score, with the cut(s) that flagged them:

The verified flagships (IDH3B, ATAD1, XYL1, psaC, COX6B1, SPR, COI1) are written up in
PDB/STRUCTURE_PAPERS.md. For human genes the structure typically sharpens an
existing IEA term or pins complex membership rather than revealing function from scratch.

Contested catalytic functions (structure adjudicates)

130 genes with a structure have a review that marked a catalytic molecular function as
REMOVE or over-annotated. A deposited structure is decisive here — it shows whether the
cofactor/active-site pocket is actually present. Two distinct cases (don't conflate them):

• Over-general parent marked over-annotated (e.g. DOT1/SIRT2 "transferase activity",
  XYL1/pobA "oxidoreductase activity"): the bound cofactor confirms catalysis and the
  structure points to the specific child term that should replace the generic one.
• Genuinely-wrong specific activity marked REMOVE (e.g. ARATH HEN1 "peptidyl-prolyl
  isomerase" — but bound SAH argues for its real RNA-methyltransferase activity; human
  CASP3 tagged "aspartic-type endopeptidase" when it is a cysteine protease; BRCA2
  "histone acetyltransferase", where the review keeps GOA's NOT annotation and leaves the
  positive H3/H4 claims UNDECIDED): the structure/cofactor adjudicates against the wrong call.

The BRCA2 row misstates the review. The current BRCA2 review does not assert
intrinsic HAT activity: it accepts GOA's negated annotation NOT|enables
GO:0004402 (histone acetyltransferase activity, IDA, PMID:9824164), which records
that the activity comes from associated P/CAF, and it marks the positive H3/H4 HAT
claims (GO:0010484/GO:0010485, IDA, PMID:9619837) UNDECIDED rather than REMOVE.
The REMOVE in pdb_gene_enriched.tsv predates the current review.

Full list with all disputed terms per gene: pdb_gene_enriched.tsv (candidate_reason
contains contested; contested_cat_mf lists the term/label/action). As always, the
disputed-term mapping and the structure PMID must both be verified before use.

Grounding in the structure papers

A bound ligand is the clue; the primary structure paper carries the functional
interpretation. PDB/STRUCTURE_PAPERS.md records verified, PubMed-sourced
notes for the shortlist above (and the prokaryotic flagships mcrA, merA, pcaF), with the
GO-annotation implication for each.

Caveat surfaced by doing this: the structure_papers PMIDs are the RCSB per-entry
primary citation — the paper that deposited that coordinate set, which is often a downstream
ligand/inhibitor or methods study rather than the definitive structure/function paper. Verified
drift cases: SPR's PMIDs are inhibitor-screening papers; cbh1's are glycosylation/propranolol
NMR; merA's is the N-terminal NmerA-domain NMR. Others (XYL1, psaC, IDH3B, ATAD1, pcaF, COX6B1)
are the definitive paper. Each PMID must be verified against the gene before it is cited in
a review (the "verify, don't trust" rule), exactly as STRUCTURE_PAPERS.md does.

Patterns worth noting: a cluster of methylotrophy / PQQ-dependent dehydrogenases
(METEA mxaI, xoxF1, mdh, mtdA, fae, PSEPK pedH, pqqB) and redox cofactor
enzymes (merA FAD/NADP, psaC/mcrA/pqqE Fe–S, DsrAB heme/siroheme) — these have
diagnostic cofactors visible in their structures, making them low-effort, high-yield review
targets.

Reproduce

# 1. offline inventory from cached UniProt + GOA (no network)
python3 projects/PDB/inventory_pdb.py
# 2. RCSB enrichment of the prioritized candidate genes (network)
python3 projects/PDB/enrich_rcsb.py
# 3. structure-paper -> GOA citation gap (offline); writes CURATION_GAP.md
python3 projects/PDB/curation_gap.py
# 4. ranked GAP_OPPORTUNITY review worklist (offline); writes GAP_WORKLIST.md
python3 projects/PDB/gap_worklist.py
# 5. H1 frontier test set: GAP_NO_EXP_CURATION genes (offline); writes data/h1_testset.tsv
python3 projects/PDB/h1_testset.py

See PDB/RESULTS.md for method detail, caveats, and the full output schema.

Does structural evidence fill annotation gaps? (PDB/H1_LEDGER.md)

H1_LEDGER.md tests whether structures fill GO gaps that traditional publications would
not. It records the GAP_OPPORTUNITY→GAP_NO_EXP_CURATION methodology correction, the
three-evidence-layer model of a structure paper (coordinates → low-information binding
terms; the paper's integrative hypothesis → informative function; sequence/EC → catalytic
identity), and the Layer-2 scoring pass. Bottom line: structures reliably supply first
experimental-grade evidence for under-curated proteins, but genuinely new informative
function is rare — throttled by subunit mismatch and GO expressivity.

Caveats

• The "no experimental MF GO" flag is computed from the cached *-goa.tsv. It flags genes
  whose GO is electronic-only in our snapshot; some are well-characterized in the
  literature. That is the point — the structure + literature can justify promoting the
  annotation, not that the function is unknown.
• A bound ligand is a clue, not proof: crystallization additives are filtered out, but a
  flagged ligand still needs reviewer judgment (substrate vs adventitious binder).
• UniProt DR PDB residue ranges drive the coverage metric; a gene present only as a short
  peptide in a partner's structure scores low coverage, correctly.

Are structure papers overlooked by curation? (PDB/CURATION_GAP.md)

PDB/curation_gap.py measures, for every deposited structure with a linked primary
publication, whether that PMID is cited in the gene's GOA REFERENCE column. Across
737 structure-paper × gene pairs (247 genes), only 15% are cited by GOA; 65%
are GAP_OPPORTUNITY (the paper predates the gene's last experimental annotation yet is
never referenced), and 174/247 genes cite zero of their structure papers. "Not cited"
means the structural study is absent from the evidence trail, not that the function is
unannotated — but it quantifies how under-used the structural literature is as a GO evidence
source. The lag boundary uses the latest experimental annotation year, since overall GOA
dates are inflated by IEA/IBA pipeline refreshes.

The reusable GOA-citation helper lives in core
(ai_gene_review.validation.goa_validator.referenced_pmids); the analysis is
project-specific.

Prioritized worklist (PDB/GAP_WORKLIST.md)

PDB/gap_worklist.py ranks the GAP_OPPORTUNITY papers by gene priority
(dark-MF / eukaryote / contested) plus the cofactor / ligand / complex richness of the
uncited structures, collapsed to one row per gene (the review unit). Top targets:
yeast PNO1, human RPS3, human BIRC5, human GCH1, human SIRT2, human MAPK1,
ARATH CRY2. Per-paper detail in PDB/data/gap_worklist.tsv.

Next steps

• [x] First eukaryotic batch reviewed with structure evidence: PICST XYL1, human IDH3B,
  human COX6B1, CHLRE psaC, ARATH COI1, human ATAD1.
• [ ] Work down GAP_WORKLIST.md from the top, citing the verified structure paper + PDB
  entry + bound cofactor as evidence; remaining shortlist includes prokaryotic
  PSEPK pcaF, METAC mcrA, PSEAI merA.
• [ ] Extend enrichment beyond the no-exp-MF set to genes with contested function
  (cross-reference CONTESTED_FUNCTION.md) where a structure could adjudicate.
• [ ] For complexes (≥2 protein entities), map partners to UniProt to support CC / complex
  membership annotations.
• [ ] Replace each peripheral RCSB auto-citation with the definitive structure/function paper
  (see the caveat above) as genes go to review.
• [ ] Consider adding a PDB-evidence field to the review schema (per ALPHAFOLD.md action items).

Slides

• Slides (Marp source: PDB-slides.md) — AI generated