ARHGAP23 (Q9P227) — curation notes

Working journal for the PAINT-campaign review. Provenance is recorded inline as
[PMID:NNNN "verbatim quote"]. Quotes marked as coming from a cached publication were
checked as verbatim substrings of publications/PMID_NNNN.md; quotes from publisher
supplementary files are reproduced by the committed script
ARHGAP23-bioinformatics/analyze_arhgap23.py, which parses them rather than transcribing.

1. What the automated provider returned, and what it missed

The affinage record (ARHGAP23-deep-research-affinage.md) passed its trust gates
(accession Q9P227 matches, human, self_evaluation_pairwise: win) and is accurate as far
as it goes, but it returned only three citations: PMID:15254754 (the 2004 in-silico
identification), PMID:38970683 (PKP4 at adherens junctions) and PMID:42006337 (the
VE-cadherin interactome). Its closing sentence is the problem:

"Beyond these junction-associated roles, the enzymatic GAP activity and domain functions
of ARHGAP23 have not been experimentally characterized in the available corpus."

That is wrong, and an independent sweep (Europe PMC full enumeration of 205 ARHGAP23
hits, NCBI gene2pubmed for GeneID 57636, plus partner/paralog-targeted searches) found
three tier-1 papers that do characterize its GAP activity, none of which names the gene
in its title — the exact failure mode this campaign has seen before:

PMID:27226243 is in NCBI's curated gene2pubmed set for GeneID 57636, so it was reachable
from the gene identifier alone.

2. The substrate question: three experiments, three partly-conflicting answers

RhoA — PMID:27226243 (Nat Commun 2016), full text cached. Loss and gain of function in
MDA-MB-231 cells:

• PMID:27226243
• PMID:27226243

Note the second clause of the first quote carefully: in these cells Arhgap21/23 knockdown
decreased active Rac1. That is the opposite of what a Rac1 GAP knockdown should do, and
it is the first half of the tension below.

Rac1 — PMID:28114311 (PLoS One 2017), full text cached. A Rac FRET biosensor in CHO.K1
cells:

• PMID:28114311
• Same paper, neurons: two independent shRNAs against rat Arhgap23 increase spine density
  and produce immature, filopodial spine morphology.

Both, in a head-to-head screen — PMID:32203420 (Nat Cell Biol 2020). Main text is
paywalled and absent from PMC, but the publisher's Source Data for Fig. 1b is open and is
parsed by the committed script. ARHGAP23's row (cellular FRET biosensor, normalised ΔR/R0,
authors' own significance flags):

The same paper's Supplementary Table 2 records the resulting call for ARHGAP23 as the
dual-specificity class RhoA+Rac1, and its literature-review columns cite only
PMID:27226243 and list nothing under "in vitro".

Synthesis. RhoA activity is supported by two independent systems (pulldown in
MDA-MB-231, biosensor in the screen). Rac1 activity is supported by two independent
systems (biosensor in CHO.K1, biosensor in the screen) but is contradicted by
PMID:27226243's Rac1 measurement in MDA-MB-231. Cdc42 is scored negative in the only
experiment that tested it. The honest reading is: RhoA well supported, Rac1 supported but
cell-type dependent, Cdc42 not supported. All of it is cellular; no assay on purified protein exists. The one systematic in-vitro
study of the family (PMID:27481945) is reported to list ARHGAP23 in its table of all 66 human
RhoGAPs without selecting it for purification — that comes from an external full-text check and
could not be confirmed from the abstract-only cache, so the independently checkable form of the
claim is the one the review leans on: Müller et al.'s Supplementary Table 2 fills the "in vivo"
and "reference" columns for ARHGAP23 and leaves the "in vitro" columns empty. That substitute
is parsed by analyze_arhgap23.py and printed in RESULTS.md §5, with the column groups read
from the header rows and a hard error if the "in vitro" block is not found — so a missing column
cannot be reported as an empty one.

3. The catalytic residue — and the published mutant that does not hit it

Committed analysis: ARHGAP23-bioinformatics/ (analyze_arhgap23.py, RESULTS.md,
results.json; --self-test plus mutation_test.py for the guards).

• The arginine finger is retained. ARHGAP1/p50RhoGAP's own annotated finger (R282,
  resolved in PDB 1TX4, the RhoA:GDP:AlF4 transition-state complex) maps onto ARHGAP23
  R942, which is ARHGAP23's own annotated Site, and the projection is reciprocal. The
  same holds against the closest paralog ARHGAP21 (R1184).
• Retention proves nothing on its own, and the controls say so quantitatively: of two
  proteins known to be GAP-dead, the residue test calls one of them "retained" (ARHGAP11B,
  which has two curated NOT|enables GO:0005096 IDA rows yet keeps its arginine); it
  correctly calls OCRL "not retained" — UniProt: "the Rho-GAP domain lacks the catalytic
  arginine and is catalytically inactive" (ECO:0000269). Family-wide, 60 of 66 human
  RhoGAP-profile proteins have an arginine at their annotated finger.
• Interface conservation does not discriminate either: ranked by identity at the 25
  1TX4 GAP:GTPase contact positions, the GAP-dead ARHGAP11B (64%) scores above the
  experimentally active ARHGAP21 (52%). That number tracks distance from ARHGAP1, not
  catalysis, and is reported as context only.
• What does carry information: at those same 25 interface positions, ARHGAP23 and the
  experimentally active ARHGAP21 carry the same residue at 22.
• The mutant discrepancy. PMID:32203420's Supplementary Table 4 describes its ARHGAP23
  imaging construct as a "GAP-deficient arginine finger mutant", and the SI figure names it
  ARHGAP23-R986K. Supplementary Table 1 gives the screened cDNA as NP_001186346.1,
  1491 aa — the same length as Q9P227, so the numbering is directly comparable. But 986 is
  not the UniProt/PROSITE arginine finger (942): it aligns to ARHGAP1 R323, which is in
  the 1TX4 interface but does not contact the nucleotide/AlF4/Mg transition state,
  whereas R282 does. That mutant was used only for TIRF localization; the arginine-finger
  controls that validated the activity screen were other proteins — [supplementary,
  machine-extracted: "mutation of critical arginine residues ('arginine fingers') abrogated
  the activity of the RhoGAPs ARHGAP11A, ARHGAP40, ARHGAP4, FAM13A, and SYDE2 (Extended
  Data Fig. 2h)."]. So no experiment has tested whether R942 is required for ARHGAP23's
  measured activity, and one of the two residue assignments is wrong.

This is reported as a discrepancy, not a correction: the structural argument favours 942,
but only an experiment settles it.

4. Localization and partners

• Lateral, non-protrusive cortex, Prickle1-dependent
  PMID:27226243 and
  PMID:27226243.
• Focal adhesions. PMID:32203420's TIRF screen places ARHGAP23 among the FA-enriched
  regulators with a pericentric distribution (Supplementary Table 4, parsed by the script:
  "Focal adhesion localization found by TIRF imaging (GAP-deficient arginine finger
  mutant), pericentric."). Independently, ARHGAP23 knockdown in CHO.K1 gives nascent
  adhesions and its overexpression gives adhesion maturation PMID:28114311.
• Adherens junctions / PKP4. PKP4-GFP affinity purification pulls down ARHGAP23 and,
  notably, not ARHGAP21 PMID:38970683. Functionally,
  PMID:38970683.
• VE-cadherin interactome PMID:42006337 — an interaction, not a localization result for ARHGAP23; the paper's
  co-IP/co-localization follow-up was on ARVCF.
• α-catenin (CTNNA1) appears as a "GOLD"-tier ARHGAP23 AP-MS interactor in
  PMID:32203420's Supplementary Table 3. This matters because a secondary source
  (PMID:33646271) states that ARHGAP21 and ARHGAP23 "have been shown to bind directly to
  α-catenin", citing two papers that on inspection assay ARHGAP21 (PMID:15793564) and an
  E-cadherin BioID list (PMID:24338363) respectively. The claim is family-level
  generalisation in that source, but the Müller AP-MS table is independent primary support
  for the ARHGAP23–CTNNA1 association.
• PDZ domain unprofiled: PMID:23722234 so the only PDZ-ligand
  claim for this protein (PTEN C-terminus, PMID:37305272) is pure docking.

5. Where the existing GOA rows come from

Eight rows, and they collapse to three distinct assertions:

• GO:0005096 × 3 — one IBA (PANTHER node PTN008592632, donor UniProtKB:Q5T5U3 = ARHGAP21)
  and two Reactome TAS. The IBA node is well grounded: ARHGAP21 carries its own enables GO:0005096 IDA from PMID:15793564. The two Reactome reactions are the RhoA set
  (R-HSA-8981637) and the Rac1 set (R-HSA-9013144), and — unlike the ARHGAP11B case in this
  batch — ARHGAP23's membership in both is explicitly sourced in Reactome's own summaries:
  "ARHGAP23 (Müller et al. 2020)" for RhoA and "ARHGAP23 (Martin Vilchez et al. 2017; Müller
  et al. 2020)" for Rac1. Martin-Vilchez et al. 2017 is PMID:28114311. So the set membership
  is correct here; what is lost is that all three rows flatten to the same substrate-blind
  term.
• GO:0005829 cytosol × 2 — the compartment attached to those same two Reactome reactions.
• GO:0007165 (InterPro IEA from the RhoGAP domain) and GO:0051056 (Reactome pathway
  membership) — both true and both strictly less informative than what the direct evidence
  supports.
• GO:0070062 extracellular exosome — HDA from a parotid-gland exosome MudPIT
  (PMID:19199708), a bulk proteomics inventory.

6. The ontology gap

Checked across three independent services, because this is load-bearing:

• GO:0005100 "Rho GTPase activator activity" is obsolete/merged: OLS4 reports
  is_obsolete: true; api.geneontology.org returns an empty record for it; QuickGO
  silently resolves the id and returns GO:0005096 instead.
• GO:0005096 "GTPase activator activity" now carries Rho GAP activity, Rho GTPase activator activity, Rac GAP activity, Rac GTPase activator activity (and the ARF,
  Rab, Ral, Ran, Rap, Ras, Sar equivalents) as narrow synonyms.
• GO:0005096 has no is_a children: OLS4 hierarchicalChildren = 0,
  api.geneontology.org subgraph descendents = 0, QuickGO returns one child, GO:1902773,
  via capable_of — not is_a.

So the molecular-function branch cannot express which GTPase a GAP acts on. The
biological-process branch still can: GO:0035024 negative regulation of Rho protein signal transduction
and GO:0035021 negative regulation of Rac protein signal transduction both exist. The
specificity that three papers measured is therefore recordable as process and not as
activity — which is the single most useful thing this review has to say to a curator.

7. Disease and other associations (background, not curated)

• 17q12 eosinophilic-esophagitis risk locus, ARHGAP23 the named gene (PMID:33446330,
  rs7211886, meta P = 5.4 × 10⁻⁷). No functional follow-up.
• 5′ partner in an ARHGAP23–FER fusion in a myofibroblastic neoplasm (PMID:34545945); the
  gene donates its promoter, which says nothing about the protein.
• PMID:38022849 is the only other ARHGAP23-titled paper; it is database mining and its
  opening sentence ("ARHGAP23 is known to activate RHO-GTPase") inverts what a GAP does.
  Not usable as functional evidence.
• No knockout phenotype exists in any organism. The nearest in-vivo handle is indirect:
  Arhgap23 is downregulated in LEC-Foxc1;Foxc2 double-knockout mouse embryos and FOXC1/2
  bind conserved regions of the locus (PMID:32510325).