ARHGAP21 provenance analyses — results

Five checks supporting genes/human/ARHGAP21/ARHGAP21-ai-review.yaml. All are
rerunnable and derive the repo root rather than hardcoding a worktree path.

uv run --with requests python resolve_partners.py      # -> partners.json
uv run --with requests python check_corrections.py     # -> corrections.json
uv run --with requests python check_gap_terms.py       # -> gap_terms.json
uv run --with requests python reference_coverage.py    # -> reference_coverage.json
uv run --with requests python intact_methods.py        # -> intact_methods.json
uv run python audit_review.py --self-test              # invariants over the review YAML

audit_review.py needs no network. The other four query UniProt, QuickGO,
NCBI E-utilities and IntAct, so their JSON outputs are committed as the record
of what those services returned.

1. check_gap_terms.py — GO cannot express GAP substrate specificity

The headline result. Each historical substrate-specific GAP term was resolved
through QuickGO's /complete endpoint, which is the only endpoint that
distinguishes a merged id from an absent one (OLS reports both
identically).

GO:0005096 has 0 is_a children (its only child is GO:1902773 GTPase activator complex, via capable_of).

Consequence. GO:0005096 is already maximal, so ARHGAP21's contested
substrate preference — Cdc42 vs RhoA vs RhoC — is not expressible in the GO
molecular-function branch. No replacement term is proposed and no new term is
requested, since the merge was deliberate. Substrate identity is instead recorded
as core_functions[].substrates and as has_input (RO:0002233) extensions.

A separate query settles the PDZ question: a QuickGO full-text search over GO
returns exactly two PDZ terms, one obsolete. The survivor, GO:0030165, is
defined "Binding to a PDZ domain of a protein" — the ligand side — and has no
descendants. GO has no term for the PDZ-domain side of a PDZ/PBM pair.

2. reference_coverage.py — coverage, not over-annotation

QuickGO queried by reference, not by gene. None of the ten primary-paper
result sets was paginated, so these counts are complete.

Seven of ten primary papers produced no GO annotation anywhere. Every paper
reporting RhoA or RhoC activity is in that set, so the GO record is silent on
that half of the literature for a reason independent of the ontology gap.

Symbol-collision check: negative. Both 2005 papers call the protein
ARHGAP10, which is this gene's former symbol and now belongs to a different gene
(A1A4S6). No annotation from any primary paper landed on A1A4S6. The seven
large-scale screens are paginated, so for those the check is reported as not
possible rather than as clean — an empty first page of 18,539 annotations is
not evidence.

3. intact_methods.py — per-partner evidence quality

502 IntAct rows, 440 distinct partners. Domain column maps IntAct's recorded
participant ranges onto UniProt's domains (PDZ 50–159, PH 931–1040, RhoGAP
1147–1339).

Two conclusions the GOA rows alone cannot give:

• Every holdup assay row maps to the PDZ domain. This mattered because
  PMID:36115835's full text never names ARHGAP21 (0 occurrences of ARHGAP in
  302 kB), so the claim had to be checked against participant ranges rather than
  assumed.
• 14-3-3 binding is genuinely replicated — 4 distinct publications and 3
  orthogonal methods for YWHAZ — not one screen logged under several sub-method
  names. This is what justifies modifying those rows to GO:0071889 while the
  other generic-binding rows are removed.

Implementation note. The first version parsed IntAct ids as db:ACC and
reported 0 partners from 502 rows, i.e. every GOA partner "ABSENT" — a
silent zero that reads as a finding. IntAct's format is ACC (db). The script
now asserts that a non-zero fraction of rows parsed and that not every GOA
partner is absent, so the same failure would abort rather than report.

4. resolve_partners.py — WITH/FROM and partner identities

15 distinct accessions: 0 dead, 0 TrEMBL, 15 reviewed Swiss-Prot, 3
non-human (mouse Arf1 P84078, HPV16 E6, HPV18 E6).

The one Xeno flag is harmless. UniProt records ARHGAP21's ARF1 interaction
against mouse P84078 rather than human P84077, which would normally mean an
IPI row asserts a cross-species interaction. Fetching both FASTA records shows
they are sequence-identical, 181 aa. Bookkeeping, not a biological defect —
recorded as a negative result.

All seven IBA rows are self-referential: WITH/FROM is PANTHER:PTN… plus
ARHGAP21's own accession Q5T5U3. That is the expected marker of a PAINT
curator judging the function core, classified NO_FAILURE_CORE, never
CIRCULAR_OR_REDUNDANT. Two nodes are involved: PTN008592632 carries
GO:0005096; PTN002754039 carries the six Golgi / actin / junction /
microtubule-transport terms.

PANTHER PTHR23175 holds 5182 proteins, of which 7 are reviewed
(Swiss-Prot) — the entries CSV is the Swiss-Prot subset, not the family.
ARHGAP21 sits in SF16 with its mouse and Xenopus orthologs; ARHGAP23
(Q9P227) is the close paralog, in SF5. ARHGAP10 (A1A4S6) is absent from the
family, so the symbol collision is not a phylogenetic relationship.

5. check_corrections.py — retractions and errata

CommentsCorrectionsList/RefType read off each cited article's own PubMed
record. This is the only way a Publisher Correction is discoverable; a
publication-type query cannot see one.

All 30 cited PMIDs returned. One flag: PMID:36115835 →
PMID:36477203, an Author Correction whose scope is "errors in Fig. 2, Fig. 4
and Fig. 5" — missing PCC values, axis labels, legends and panel ordering.
Cosmetic; no data changed, so the interactions it reports stand. No
retractions.

Mutation-tested by inserting PMID:32125225, a known retracted paper, which the
checker flagged retracted_by_pubtype=True with RetractionIn: PMID:35078223.

6. audit_review.py — invariants over the review YAML

Current state: 47 GOA data rows → 47 non-NEW entries + 4 NEW = 51, and 50
supporting_text quotes verified verbatim under whitespace normalisation.

Checks, each covering something the repo validator does not:

1. Row reconciliation against the GOA TSV by (term, evidence, reference, with/from) multiset. The fetch-gene stub is known to collapse distinct
   GO:0005515 partner rows; here it did not, but the count is asserted rather
   than eyeballed.
2. Duplicate YAML keys, via a SafeLoader subclass that raises on repeats.
   PyYAML keeps the last occurrence and discards the earlier one silently, and
   every other gate in this repo walks the parsed document — so data destroyed
   by parsing cannot fail them.
3. Quote verbatimness for both PMID: and file: references. CI checks only
   the former.
4. No row left PENDING, no surviving TODO, status: COMPLETE.
5. Reference completeness — every PMID: cited in the document resolves to a
   references entry. The repo validator only checks original_reference_id,
   so a PMID named in a reason, a review_notes, a knowledge_gap or the
   description can otherwise dangle. Added in round 3, after rounds 1–2 had
   asserted this invariant in prose while it was enforced only by a throwaway
   script outside the repo.

Writing it went wrong twice, both caught by the self-test, not by reading:
the first version scanned the references block too, so deleting an entry
deleted the only occurrence of its PMID and the guard stayed silent
(a check partly satisfied by its own subject reports success); excluding
the whole block then over-corrected, losing PMID:36477203, which is cited
only inside PMID:36115835's review_notes — the exact case the check
existed for. The working scope strips only the reference id fields. The
cited-PMID count read 24 → 22 → 23 across the three scopes, and the
middle figure looked perfectly reasonable.

--self-test mutates an in-memory copy for each check and requires the check to
fire; every mutation asserts its anchor exists first, so a drifted anchor is an
error rather than a false pass. A passing self-test proves the guards that exist
work — it cannot say which guard was never written.

What no script here can check

None of these gates validates a quote against the claim it is attached to.
Two rows in this review initially carried a verbatim, correctly-attributed quote
from PMID:15161933 describing that paper's 14-3-3 affinity-column method,
sitting under summaries about BioPlex AP-MS. Every mechanical check passed. It
was found by re-reading each summary next to its quote, and fixed by quoting
each row's own paper.

Review round 2 found four more of the same class, which is the strongest
available evidence that this gap is structural rather than a one-off:

• GO:0032956's summary asserted a stress-fibre phenotype that appears nowhere
  in the cited abstract (it is a UniProt MISCELLANEOUS line, and one that
  spans a CC continuation so it cannot be quoted verbatim at all).
• The same row's reason justified its evidence code using the other NEW
  row's two-hybrid partner.
• GO:0051684 maintenance of Golgi location was quoted with the
  nocodazole-washout sentence, which is establishment evidence.
• GO:0005794 IDA was quoted with an introduction sentence ending in reference
  callouts "(23, 24)" — prior work, not the paper's own observation.

All four passed every gate here: real PMIDs, verbatim strings, correct terms,
correct actions. Two cheap heuristics came out of it and are worth applying
before attaching any quote to an experimental row: a sentence containing
reference callouts is background, not result, and check that the assay in
the quote is the assay the term names (re-establishment after dispersal is not
maintenance).