ARHGAP4 (P98171) — review notes

X-linked RhoGAP at Xq28, formerly C1 / RGC1 / p115 / KIAA0131. 946 aa, with
F-BAR (19–317), Rho-GAP (507–695) and SH3 (746–805) domains
[file:human/ARHGAP4/ARHGAP4-uniprot.txt "FT DOMAIN 507..695"]. Immediately
adjacent to AVPR2, which is why almost all human genetic information about it comes
from contiguous-gene deletions.

Headline: this is a coverage problem, not an over-annotation problem

reference_coverage.py queried QuickGO by reference for 19 primary ARHGAP4
papers. 16 of 19 produced no GO annotation on ARHGAP4 in any species. The three
that did:

Human P98171 therefore has no experimental molecular-function or biological-process
annotation at all. Its only IDA rows are localisation (HPA cytosol, LIFEdb cytoplasm).
Everything functional is IBA, Ensembl-projected IEA, or 1996 TAS.

The Ensembl GO_REF:0000107 rows are not free-floating: every one projects a curated
rat row, and the rat rows trace to the two mechanistic papers above. So the human
record is a faithful shadow of rat experiments — which is the right thing for it to be,
and worth stating rather than second-guessing.

Substrate specificity, and why GO cannot hold it

The only direct activity measurement is rat: "In vitro, recombinant ARHGAP4 stimulated
the GTPase activity of three members of Rho GTPases, Rac1, Cdc42 and RhoA"
PMID:12414125.
Reactome commits further, filing ARHGAP4 under R-HSA-9013144 RAC1 GAPs stimulate RAC1 GTPase activity.

Neither commitment can be written as a GO term. check_gap_terms.py resolved the
candidate terms across three independent services (OLS4, the GO API at
api.geneontology.org, and this repo's cache/go/terms.csv), because QuickGO
silently resolves merges — ask it for GO:0005100 and it returns GO:0005096's
record flagged isObsolete: false, which reads as "the term is fine". It did exactly
that for 7 of the 9 ids queried. Across the independent services:

• GO:0005100 Rho, GO:0005099/GO:0005098 Ras, GO:0008060 ARF, GO:0005097 Rab,
  GO:0046582 Rap GAP activity — all merged into GO:0005096;
• GO:0017048 Rho GTPase binding — merged into GO:0031267 small GTPase binding, so
  the binding route loses Rho specificity too;
• GO:0005096 has zero is_a children. It is not childless: QuickGO reports one
  child, GO:1902773 GTPase activator complex, via capable_of — a complex that is
  capable of the activity, not a more specific activity. OLS4 reports zero direct
  subclasses.
• the merged names survive only as synonyms of GO:0005096 ("Rho GTPase activator
  activity", "Rac GAP activity", …).

So the merge was deliberate and complete, and no term is being requested. Substrate
identity is carried in core_functions[].substrates and as RO:0002233 has_input
extensions instead. This is the same conclusion the sibling ARHGAP21 review reached, on
the same terms, reached here independently and with the QuickGO caveat added.

The arginine finger: present, and that settles nothing

arginine_finger.py aligns UniProt-delimited Rho-GAP domains to ARHGAP1/p50RhoGAP and
projects the structurally resolved catalytic arginine.

The anchor is verified structurally, not assumed: PDB 1TX4 is the
RHOA·GDP·AlF4⁻·p50RhoGAP transition-state complex; SIFTS maps chain A 1–198 onto
Q07960 234–431, so Arg-282 is deposited residue 49 — and the script confirms residue
49 is ARG with its guanidinium 2.73 Å from the AlF4⁻ moiety.

Result: ARHGAP4 carries R543 at that column. The computed alignment agrees with
UniProt's independent PROSITE-ProRule SITE in all 11 panel members, positives and
known-dead alike (ARHGAP36 T258, DEPDC1B I231, OCRL1 Q757, INPP5B Q852 — the same four
Amin et al. 2016 PMID:27481945 list as lacking the finger).

But residue identity and curated activity are decoupled in both directions, and the
script demonstrates both from GOA rather than asserting it:

• arginine present, activity denied — ARHGAP11B (Q3KRB8) retains the arginine at
  the aligned column and carries two NOT|enables GO:0005096 IDA rows
  (PMID:25721503, PMID:27957544). UniProt names it "Inactive Rho GTPase-activating
  protein 11B".
• arginine absent, activity annotated — OCRL1 carries GO:0005096 by IDA,
  ARHGAP36 by IBA, ARAP2 by IEA+IBA, FAM13B by TAS, all without the arginine.

Amin et al. add the mechanistic reason not to over-read the residue:
PMID:27481945
and
PMID:27481945.
ARHGAP4 (listed as "p115", P98171, entry 42) appears in that paper's Table 1 of 66
RHOGAPs, inside the 57 that "have a common catalytic domain capable of terminating RHO
protein signaling", but it is not among the 14 whose kinetics were measured
PMID:27481945.
Verified by grepping the PMC full text, where the exact token ARHGAP4 occurs once, in
that table. A web-search summary claiming Amin measured an "ARHGAP4-R543K mutant by
FRET" is not in the paper and was not used.

So: the residue is intact, the fold is intact, the only activity measurement is a rat
in-vitro assay, and nothing here licenses either "therefore active in human" or
"therefore inactive".

Loss of function: unremarkable in both species

• Mouse. IMPC recorded zero significant genotype-to-phenotype associations for
  Arhgap4 across 413 statistical tests. The second number is the one that matters:
  an untested gene returns the same zero as a clean one, and in the same query
  Srgap1 and Srgap3 come back 0/0 — never phenotyped — while Lepr (105/244),
  Dmd (23/842) and Trp53 (6/653) confirm the query shape. So this is a tested
  absence, checkable by impc_phenotypes.py.
• Human. Contiguous Xq28 deletions removing AVPR2 plus all or part of ARHGAP4
  cause nephrogenic diabetes insipidus attributable to AVPR2
  PMID:10425039, PMID:11754100, PMID:22965914. The decisive one is Schöneberg
  1999: a boy hemizygous for a 21.5 kb deletion removing AVPR2 and most of C1/ARHGAP4
  was followed for 16 years with targeted immunologic and hematologic testing — the
  tissue where the gene is enriched — and nothing beyond the AVPR2 phenotype was found
  PMID:10425039,
  leading to
  PMID:10425039.
  The one SCID case with a 34.4 kb deletion is explicitly not attributed to ARHGAP4:
  PMID:16781893,
  and the authors attribute the SCID to loss of a conserved intergenic regulatory
  element between ARHGAP4 and ARD1A.

The compensation hypothesis has been tested once, and half-refuted. Fujimoto 2008
profiled PBMC from patients lacking about half of ARHGAP4 and found
PMID:18489790,
concluding that loss "is not compensated for by other family members". That refutes the
transcriptional version of Schöneberg's hypothesis and leaves protein-level redundancy
open — which is the version a combinatorial knockout would test, and which nobody has
done. Both papers are cited in the review's knowledge gap rather than only the one that
fits. Huang 2012 states the residual position plainly:
PMID:22965914.

Two single-family variant reports exist — T491M in an intellectual-disability family
PMID:26707211 and variants in X-linked early-onset temporal lobe epilepsy
[PMID:39060771, a Letter with no abstract in the PubMed record]. Neither produced a GO
annotation and neither is, on its own, evidence of a molecular function. Note that
T491 lies outside the Rho-GAP domain (507–695).

PANTHER placement and what the IBAs actually assert

ARHGAP4 sits in PTHR14166 "SLIT-ROBO Rho GTPase-activating protein" alongside
SRGAP1/2/3 and the hominin-specific SRGAP2B/2C. resolve_entities.py joins each
WITH/FROM token onto the family's PAINT export:

Three things follow.

1. PTN002680572 is seeded by UniProtKB:P98171 alone. Per CLAUDE.md this is the
   legitimate marker that experimental grounding exists on the target (ARHGAP4 has
   cytoplasm IDA from LIFEdb), not circularity.
2. GO:0051963 regulation of synapse assembly is the one inherited assertion whose seed
   set contains no ARHGAP4-lineage member — it is Srgap2, Srgap3, and two
   hominin-specific paralogs that arose by partial duplication of SRGAP2.
3. The same curator placed the synaptic localisation terms on a sibling node,
   PTN002680580, which ARHGAP4 does not inherit. The boundary excluding ARHGAP4
   from the synaptic compartment was therefore drawn deliberately, which makes the
   synapse-assembly inheritance the odd one out rather than part of a consistent
   neuronal package.

The one REMOVE, and why it took a review to get right

GO:0005515 protein binding IPI with NCKAP1L (Hem-1, P55160) — UniProt's only SUBUNIT
statement for this gene, from the neutrophil leading-edge complex proteomics
PMID:16417406.

My first pass marked it KEEP_AS_NON_CORE, reasoning that the interaction is real and the
partner informative even if the term is not. That is not a branch the repo policy offers.
.claude/skills/annotation-reviewer/SKILL.md:186-197 gives exactly three: MODIFY when
the paper supports a more informative MF, UNDECIDED when the evidence cannot be
adjudicated, and otherwise REMOVE — "removal does not mean the reported interaction
is false". UNDECIDED does not apply here, because the interaction is curated with
experimental evidence and is not in doubt; what is in doubt is whether any informative MF
term exists, and the branch walk answers that: no.

So the row is REMOVE. The useful content — that a hematopoietic RhoGAP sits inside the
machinery that activates RAC — survives in the review prose, the description and a
suggested_questions entry, which is where it belongs, rather than in an annotation that
says only "binds something".

Is there a term to move to? Answered by walking the branch, not by searching, because
GO search is token-based and "WAVE complex binding" can never match a term lacking those
tokens. A complete enumeration — unbounded in depth, deduplicated — returns 105
distinct descendants of GO:0044877 protein-containing complex binding, reached over
106 parent-child edges with every descendant within four edges of the root,
containing exactly one actin-machinery term, GO:0071933 Arp2/3 complex binding. GO:0031209 SCAR complex exists but is a cellular component, not something to bind — and in any case the
paper describes the ARHGAP4-containing complexes as biochemically separable from the
WAVE2 complex proper, so SCAR-complex membership would be wrong on the biology as well as
unavailable in the ontology.

Comparator check — which refuted three of my four candidate MODIFYs

comparator_terms.py tests each proposed replacement as a prediction about seven
well-curated RhoGAPs (ARHGAP1, ARHGAP35, ARHGAP17, ARHGAP21, ARHGAP24, SRGAP2, DLC1):

GO:0007266 on a RhoGAP is an established convention (ARHGAP1 TAS, ARHGAP35 IBA), not
an error. GO:0007165 is a uniform InterPro2GO product on all seven comparators, so
rewriting it on one gene would be idiosyncratic. No comparator carries cell cortex,
despite the 1996 paper describing "a narrow cytoplasmic region just below the plasma
membrane" — so that observation stays a knowledge gap, not an annotation.
GO:0007010 cytoskeleton organization is the genuine outlier: no comparator carries
it, and the cited observation is inhibition, which is a regulation term.

Reproducing the analyses

cd ARHGAP4-bioinformatics
uv run --with biopython --with requests python arginine_finger.py     # -> arginine_finger.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 resolve_entities.py                     # -> entities.json
uv run --with requests python comparator_terms.py                     # -> comparator_terms.json
uv run --with requests python impc_phenotypes.py                      # -> impc_phenotypes.json
uv run --with pyyaml python audit_review.py                           # invariants, no network

Each script takes --self-test, which breaks the document or the anchor on purpose and
requires every guard to fire, with negative controls that must stay silent.

Reference caveat

Affinage cites PMID:40817404 (NPJ Precis Oncol 2025) for the MYH9/β-catenin/c-Jun
result. An earlier version of the same work, PMID:40553870, is WITHDRAWN from
Int J Biol Macromol — "withdrawn at the request of the author(s) and/or editor due to
an error in the publishing process", i.e. a publishing error, not a data retraction. The
republished version stands, but the pair is worth flagging.

Affinage's record passed its trust gates (accession P98171, no non-human organism
token, self_evaluation_pairwise: win, faith_pct: 100) and its narrative is accurate
as far as it goes, but it returned 11 citations and missed every human-genetics
paper (PMID:16781893, PMID:18489790, PMID:10425039, PMID:22965914, PMID:26707211,
PMID:39060771) and the family-wide RHOGAP kinetics survey (PMID:27481945). Those were
found by PubMed queries on the symbol and its synonyms, not by the provider. The pattern
is the one this campaign has seen before: every missed paper is titled for the disease
(nephrogenic diabetes insipidus, SCID, intellectual disability, epilepsy) or for the
family rather than for ARHGAP4's function, so a function-shaped retrieval never
surfaces them. Two of them — Schöneberg 1999 and Fujimoto 2008 — turn out to carry the
single most load-bearing evidence in this review.