ARHGAP21 (Q5T5U3) — review notes

Human Rho GTPase-activating protein 21, 1958 aa, HGNC:23725, chromosome 10p12.

1. Nomenclature: this gene used to be called ARHGAP10, and ARHGAP10 is now a different gene

The two founding papers, and several later ones, call this protein ARHGAP10
PMID:12056806.
UniProt still lists ARHGAP10 as a synonym of ARHGAP21. HGNC has since assigned the
symbol ARHGAP10 to a different gene (UniProt A1A4S6, a GRAF/oligophrenin-family
protein). The later literature flags the collision explicitly
PMID:20525016.

Checked, and clean. reference_coverage.py queries QuickGO by reference for every
primary paper and asks whether any annotation landed on A1A4S6. None did. The
PMID:15793564 and PMID:16184169 annotations sit on Q5T5U3 (and, for the latter, on
P35221 CTNNA1), which is correct. Recording the negative so the next reviewer knows the
check was run rather than skipped.

ARHGAP10/A1A4S6 is also not a member of ARHGAP21's PANTHER family (see §6), so it is
not a paralog in the phylogenetic sense either — only a symbol collision.

2. Domain architecture — there is no ArfGAP domain

UniProt's feature table gives exactly three domains
[file:human/ARHGAP21/ARHGAP21-uniprot.txt "FT DOMAIN 1147..1339"]:

plus two mapped interaction regions: 930–1097 "Interaction with ARF1 and ARF6"
(ECO:0000269|PubMed:15793564) and 1592–1861 "Interaction with CTNNA1"
(ECO:0000269|PubMed:16184169).

The 930–1097 region is an ARF-binding domain (the authors' "ArfBD"), not an ArfGAP
domain. The crystal structure resolves it as a PH domain plus a C-terminal helix
PMID:17347647.
An independent ARF6 study classifies ARHGAP21 as an effector, explicitly contrasting it
with the Arf GAPs in the same experiment
PMID:16527809.

So ARHGAP21 is a Rho-family GAP that is an ARF effector, not a dual-specificity
Rho/ARF GAP. This matters because the name-implied-catalysis error (a GAP term inferred
from a domain or region name that was never assayed) is the commonest defect class in this
campaign — here GOA has not made it: there is no ARF-GAP annotation on the gene.

3. The catalytic machinery is not in dispute

The arginine finger is R1184, and UniProt records the mutagenesis as
R->A: Loss of GTPase activity and loss of function. with evidence from both founding
papers (ECO:0000269|PubMed:15793564, ECO:0000269|PubMed:16184169). Both groups
independently built the same catalytically-dead mutant and both saw loss of function.

That is the key framing for §4: the disagreement is about which GTPase, never about
whether ARHGAP21 is a GAP.

4. Substrate specificity is genuinely contested — and GO cannot express the disagreement

4a. What the papers actually say

Verbatim anchors:
PMID:15793564,
PMID:16184169,
PMID:23200924,
PMID:23235160.

The dispute is narrower than "RhoA vs Cdc42". Both 2005 papers agree ARHGAP21 has
Cdc42 GAP activity; Dubois says Cdc42 is preferred (not exclusive), Sousa says RhoA is
also a substrate. What is genuinely unresolved is (i) the rank order of preference and
(ii) whether the RhoA/RhoC activity is constitutive or context-restricted. Lazarini's own
data argue for context: the same protein showed the activity in PC3 but the phenotype was
absent in LNCaP. UniProt records both without adjudicating
[file:human/ARHGAP21/ARHGAP21-uniprot.txt "CC -!- FUNCTION: Functions as a GTPase-activating protein (GAP) for RHOA and"].

I have not collapsed this into one verdict, and the review should not either.

4b. GO cannot record which GTPase — measured, not assumed

check_gap_terms.py resolves each historical substrate-specific GAP term through QuickGO's
/complete endpoint (the only endpoint that distinguishes a merged id from an absent
one — OLS reports both identically):

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

So GO:0005096 is already maximal: there is no more specific term to move to, and
per the campaign's standing rule I am not filing a proposed_new_terms entry to
re-create what GO deliberately merged. The consequence is recorded as a KnowledgeGap
instead: the substrate identity of a GAP is not expressible in the GO molecular-function
branch at all. The machine-readable alternative is core_functions[].substrates plus a
has_input (RO:0002233) extension on the annotation, which is what this review uses.

The same merge applies on the binding side: GO:0017137 (Rab) and GO:0017048 (Rho)
GTPase binding both resolve to GO:0031267 small GTPase binding.

4c. The GO record is silent on RhoA/RhoC for a second, independent reason

reference_coverage.py (QuickGO by reference) shows that every paper reporting RhoA or
RhoC activity — PMID:16184169, PMID:21173159, PMID:23200924, PMID:29212046 —
produced zero GO annotations anywhere in GOA. So even setting the ontology gap aside,
the RhoA/RhoC half of the literature was never curated.

Note the precision this needs: because GO:0005096 is substrate-agnostic, GOA does not
assert "Cdc42" either. GO says only "this is a GAP". The disagreement is invisible in GO
twice over — once because the term cannot carry a substrate, and once because the papers on
one side of it were never read into GOA.

5. The PDZ domain is a validated PDZ-binding-motif receptor, and GO has no term for that side

Ten of the 24 GO:0005515 rows come from one paper, PMID:36115835 (Gogl 2022), a
systematic PDZ-domain / PDZ-binding-motif affinity survey
PMID:36115835.
Its full text never names ARHGAP21 (0 occurrences of the string ARHGAP in 302 kB), so
"these rows measure the PDZ domain" had to be checked rather than assumed.

intact_methods.py checks it against IntAct's recorded participant ranges: all ten
holdup assay rows map onto residues 50–159, i.e. the PDZ domain. Five of the ten
(AXIN1, APC, CTNNB1, MYO18A, SAPCD1) additionally carry an orthogonal anti tag coip row
from a different publication, so they are not single-method hits.

But there is no GO term for this. A QuickGO full-text search over GO returns exactly two
PDZ terms, one of them obsolete; the survivor is GO:0030165, defined as
"Binding to a PDZ domain of a protein" — the ligand side, for proteins that carry a
PBM and bind someone else's PDZ. It has no descendants. ARHGAP21 is on the opposite side of
that interaction, and GO offers nothing for it.

This is why these ten rows cannot be upgraded to something informative and are removed as
uninformative generic binding rather than modified. Removal does not mean the interactions
are false — recorded as a second KnowledgeGap.

5a. Correction (round 2): the PDZ domain is NOT functionally dark

My first draft tagged this gap dark_aspect: MF_DARK and called the PDZ "the one module of
the protein with no assigned cellular role". That was wrong, and it was refuted by a file
this same PR adds. publications/PMID_41957357.md (2026, Cell Death Discovery, human HCC
cells) reports
PMID:41957357.

So the module has a reported cellular ligand — filamin A. Two consequences:

1. The MF_DARK tag and the "no assigned role" phrasing are removed. The ontology gap
   survives untouched: finding a ligand does not create a term for the PDZ-domain side.
2. GO:0031005 filamin binding exists and is partner-specific, so unlike the ten holdup
   rows this one interaction can be expressed — added as a fourth NEW row. Note what that
   implies about the gap: GO can record ARHGAP21's PDZ interactions only where it happens to
   have a partner-specific binding term, which is a ligand-by-ligand workaround that does not
   scale to a 65,000-interaction survey.

The row is scoped to the binding, which the paper states outright, not to PDZ-dependence,
which the paper hedges as a "potential mediator". The HCC metastasis and FLNA-stabilisation
phenotypes are not proposed as process terms: they are effects of the interaction in a cancer
context, not a demonstrated normal function.

Process note worth keeping: affinage did return this paper, and I read it when checking
retractions — but I used it only for the retraction sweep and never asked what it said about
the domains. A paper can be in your own commit and still be unread for the claim it refutes.

6. Provenance checks

6a. All seven IBA rows are self-referential — which is correct, not circular

Every IBA row's WITH/FROM is PANTHER:PTN…|UniProtKB:Q5T5U3, i.e. the PANTHER ancestral
node plus ARHGAP21's own accession. Per the PAINT convention this marks a curator
judging the function core on the strength of the target's own experimental annotation; it
is NO_FAILURE_CORE, never CIRCULAR_OR_REDUNDANT.

Two distinct nodes are involved:

• PTN008592632 → GO:0005096 (the GAP activity)
• PTN002754039 → the six Golgi / actin / junction / microtubule-transport terms

6b. PANTHER family

PTHR23175 ("PDZ domain-containing protein"), 5182 proteins, of which 7 are
reviewed (Swiss-Prot) — the row count of the cached entries CSV is the Swiss-Prot subset,
not the family. Those seven are ARHGAP21 (human, mouse, X. tropicalis, two X. laevis)
in subfamily SF16, and ARHGAP23 (human Q9P227, mouse Q69ZH9) in SF5.

So ARHGAP23 is the genuine close paralog and it sits in a different subfamily.
ARHGAP10 (A1A4S6) is absent from the family entirely (§1).

6c. The one non-human WITH/FROM accession is harmless

resolve_partners.py resolves all 15 distinct WITH/FROM accessions: 0 dead, 0 TrEMBL, 15
reviewed Swiss-Prot, 3 non-human (mouse Arf1 P84078; HPV16 and HPV18 E6).

UniProt flags P84078 Xeno, which would normally mean an IPI row asserts a
cross-species interaction. Here it does not matter: human ARF1 (P84077) and mouse Arf1
(P84078) are sequence-identical, both 181 aa (checked by fetching both FASTA records
and comparing). So the accession choice is bookkeeping, not a biological defect. Negative
result, recorded.

6d. Retractions and errata

check_corrections.py reads CommentsCorrectionsList/RefType off each cited article's own
PubMed record — the only way a Publisher Correction is discoverable, since 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" consisting of missing PCC values, axis labels,
legends and panel ordering. Cosmetic; no data changed, so the ARHGAP21 interactions it
reports stand. No retractions anywhere.

The checker was mutation-tested by inserting PMID:32125225 (a known retracted paper),
which it correctly flagged retracted_by_pubtype=True with RetractionIn: PMID:35078223.

6e. Interaction evidence quality per partner

From intact_methods.py (502 IntAct rows, 440 distinct partners):

• YWHAZ — 5 rows, 4 distinct publications, 3 orthogonal methods (anti tag coip,
  proximity-dependent biotin identification, pull down). This is real replication, not
  one screen logged several ways.
• SFN — 4 rows, 2 publications, BioID + TAP.
• ARF1 — 3 rows, methods cosedimentation and x-ray diffraction, mapped to the
  PH domain. That is the 2.1 Å structure; the strongest single interaction datum on the
  gene.
• The ten PDZ holdup partners — see §5.
• EEF1D — display technology + tap only, no domain mapping; the weakest row.

14-3-3 binding across four independent papers is consistent with the gene's heavy
phosphorylation (UniProt lists ~20 phosphoserine sites), and is the one generic-binding
cluster that supports a specific replacement term, GO:0071889 14-3-3 protein binding.

7. Coverage, not over-annotation, is this gene's problem

reference_coverage.py, per primary paper (none of the ten is truncated, so these counts
are complete):

Seven of ten primary papers produced no GO annotation anywhere. ARHGAP21's entire GO
record rests on three papers, and a 2.1 Å crystal structure is recorded as bare
protein binding.

The single largest gap is PMID:16184169 — a Nature Cell Biology paper whose title
finding is that ARHGAP21 is required for α-catenin recruitment at adherens junctions. It
produced two annotations, both on α-catenin, none on ARHGAP21, even though UniProt
carries the interaction with a mapped region (1592–1861) and states it in FUNCTION and
SUBUNIT. Hence three NEW rows in this review:

• GO:0045294 alpha-catenin binding (IPI) — verified active, defined
  "Binding to catenin complex alpha subunit."
• GO:0034333 adherens junction assembly (IMP)
• GO:0032956 regulation of actin cytoskeleton organization (IMP) — the Dubois
  Arp2/3-and-F-actin finding likewise has no BP term on the gene.
• GO:0031005 filamin binding (IPI, added round 2) — see §5a.

Affinage's record (gates_passed: true) returned 15 citations and missed both founding
2005 papers and Hehnly 2010 — i.e. it missed every paper that actually carries a GOA row,
plus the α-catenin paper that drives the largest gap here. Consistent with the campaign
finding that the gates certify precision and say nothing about recall; the α-catenin paper
is titled for the partner and for Listeria, which is exactly the shape that gets missed.

6f. Sibling cross-check on GO:0005096 — no divergence

The campaign has seen three independently-reviewed paralogs give three different
answers to a byte-identical row, so sibling reviews are checked before finalising.
The repository has no other ARHGAP review, but it does hold three reviewed GAPs of
a different family (ACAP1, ACAP2, ACAP3) carrying seven GO:0005096 rows between
them. All seven are ACCEPT, which is what this review does with its three.
No inconsistency to flag — recorded because a negative cross-check is still a
result.

8. Things I could not resolve

• Rank order of GAP preference. No study has assayed Cdc42, RhoA and RhoC side by side
  with the same purified protein and the same assay. The 2005 in-vitro comparisons used
  different constructs and readouts.
• Whether the RhoA/RhoC activity is context-restricted. Lazarini's PC3-vs-LNCaP split
  is suggestive but is a single paper with a cell-line comparison, not a mechanism.
• The PDZ ligand repertoire in cells. The holdup survey gives affinities for isolated
  fragments, and exactly one cellular ligand has been reported for the domain — filamin A,
  in HCC cells, with the domain mapping hedged as a "potential mediator" (§5a). Which of
  the ten holdup partners the PDZ actually engages in vivo, and in which compartment,
  remains untested.
• The direction of the F-actin effect. UniProt records that depletion causes
  accumulation of F-actin stress fibres, while antisense knockdown in neonatal mouse islets
  reduces F-actin PMID:25744409.
  These may simply be different cell types, but nothing reconciles them, which is one reason
  the proposed BP term is the sign-agnostic GO:0032956 rather than a directional one.
• The cell system of the α-catenin knockdowns. PMID:16184169 and PMID:23235160 are
  both abstract-only in the cache, neither abstract names a cell line, and PMID:16184169
  is absent from PMC and Europe PMC full text with no text-mined annotations. An
  experimental evidence code asserts the experiment was done in the annotated organism, so
  this matters: the two IMP NEW rows keep that code because the two-hybrid partner is
  human α-catenin and the target is human ARHGAP21, but both rows now carry an explicit
  caveat asking a curator with full-text access to confirm the cell line. An earlier draft
  asserted "in human cells" in the summary; that claim could not be verified and was removed.
• GO:0051645 Golgi localization appears in UniProt's GO cross-references as
  IBA:GO_Central but is absent from the GOA TSV, which instead carries GO:0051683 and
  GO:0051684. Not acted on — it is a cross-reference snapshot difference, not an
  annotation this review can change.

8a. Round-2 evidence-attachment corrections

Six review items, all of the same class: the term and the action were right in every
case, and the quote attached to them was not. Recorded because the class is invisible to
every mechanical gate in this repo — CI checks that a supporting_text is a verbatim
substring of its source, never that it supports the sentence it sits under.

Two things worth carrying forward. First, "cites a reference callout" is a cheap, reliable
tell that a sentence is background rather than result — worth grepping for before attaching
any quote to an IDA/IMP row. Second, the reviewer's list was not perfect and was checked
rather than applied: it missed PMID:36477203 (named in the review, no reference entry), and
two of the PMIDs it listed as "omitted" were not actually named in the document until this
round added them. Every item was verified against the cached files before being acted on.

8b. Round 3: I asserted a guard existed, in prose, while it lived outside the repo

Round 2 wrote, in two places (§8a's table and the history record), that the fix script
"now asserts that every PMID string appearing anywhere in the document resolves to a
reference entry, so this cannot silently regress."

That was false. The assertion was in fix6_arhgap21.py, a one-shot scratchpad script that
is not committed and never runs again. Nothing in the repository enforced it. The invariant
happened to hold, by hand, on that one day.

This is the review's own thesis turned on its author: a true-sounding claim attached to the
wrong object. And it is the sharper variant, because the claim was about a guard — the
exact thing this gene's analysis is otherwise careful about. audit_review.py's own docstring
already warned about it: "a passing self-test proves the guards that exist work; it cannot
tell you which guard was never written."

Fixed by writing the guard rather than softening the sentence, since the invariant is
worth having: check_references is now the fifth check in the committed
audit_review.py, with a --self-test mutation.

Writing it went wrong twice, both caught by the self-test

1. The detector's scope included the lookup table it validates against. The first version
   scanned yaml.dump(doc) whole — including references. So deleting a reference entry
   removed the only occurrence of its PMID, named shrank in step with ref_ids, and the
   guard stayed silent. The self-test reported dropping reference PMID:12056806 was not detected. A check that is partly satisfied by its own subject is worse than no check.
2. The obvious correction over-corrected. Excluding the whole references block fixed the
   self-satisfaction but lost a real case: PMID:36477203, the Gogl erratum, is cited only
   inside PMID:36115835's review_notes. Under that scope it stopped counting as cited, so
   dropping its entry would not have fired — and that is precisely the citation the round-1
   reviewer had missed.

The working scope strips exactly the id fields of references and keeps everything else,
which was verified directly rather than assumed: deleting PMID:36477203's entry now
produces PMID:36477203 is cited in the document but has no references entry.

Cited-PMID count moved 24 → 22 → 23 across those three scopes. The middle number was the
over-correction, and it looked perfectly reasonable.

9. Scripts

All committed under ARHGAP21-bioinformatics/ and rerunnable
(uv run --with requests python <script>):