ARHGAP11B — review notes

UniProt Q3KRB8 (RHGBB_HUMAN), 267 aa, PE 1: Evidence at protein level.
GOA TSV: 17 rows. Seeded existing_annotations: 17. Counts reconcile exactly, so no
GO:0005515 partner rows and no same-term/different-assigner rows were collapsed by the
stub on this gene. (Checked because the stub is known to under-seed on other genes.)

What the gene is

ARHGAP11B arose on the human lineage by partial duplication of ARHGAP11A
(Q6P4F7, 1023 aa), then acquired a single C→G substitution that creates a novel splice
donor site. 55 nt are spliced out, the reading frame shifts, and the protein ends at 267
residues with a C-terminus found in no other protein.

"The first 220 amino acids of ARHGAP11B are 98% identical to those of ARHGAP11A (216 of
220 residues; for a discussion of the four amino acid changes, see below and the
Supplementary Materials) and encompass most of its GAP domain (amino acids 46 to 246)"
PMID:27957544

"Thus, the single C→G base change in the ARHGAP11B gene is ultimately the reason for the
reading frame shift that leads to the truncation of the GAP domain of ARHGAP11B and its
novel C-terminal sequence (amino acids 221 to 267; Fig. 1 and fig. S2), which is unique
to ARHGAP11B and not found in any other protein thus far described in the animal
kingdom" PMID:27957544

The substitution predates the split from Neanderthals and Denisovans and is fixed in
present-day humans [PMID:27957544, "The crucial C→G base substitution was also found in
Neanderthal and Denisova ARHGAP11B (fig. S3). Moreover, all present-day humans analyzed
(31) carry the C→G substitution."].

The central curation problem: GOA asserts GAP activity and its negation simultaneously

The GOA TSV carries both of these for the same gene:

Two independent experimental papers say the protein has no GAP activity; Reactome says it
does. This is not a subtle over-annotation — it is a flat contradiction inside a single
gene's GOA record, and the positive side is the one with the weaker evidence code.

Where the wrong positive comes from, measured

Querying QuickGO by reference rather than by gene settles it (numberOfHits compared
against len(results) each time, and paged until they matched):

R-HSA-8981637 is a Reaction in the cytosol/plasma-membrane compartments whose single
catalystActivity is a set-level entity, "GTPase activator activity of RHOA GAPs [cytosol]". So both the activity term and the cytosol term are handed to every member of
a defined catalyst set at once. ARHGAP11B is in that set because it looks like ARHGAP11A.
For ARHGAP11A the same projection is correct and should stay.

The paralog comparison, which is the cleanest evidence in the whole review

QuickGO for the two genes side by side:

The same paper produced the positive annotation on ARHGAP11A and the negated one on
ARHGAP11B, and the same study that put ARHGAP11B in mitochondria put ARHGAP11A in the
nucleus. The UniProt-sourced curation is internally consistent and correct for both
paralogs. Every wrong row on ARHGAP11B is non-experimental, and the shared Reactome cytosol
call describes neither paralog's measured location.

UniProt has already adjudicated it, twice over. The recommended name is a verdict:

"RecName: Full=Inactive Rho GTPase-activating protein 11B"
[file:human/ARHGAP11B/ARHGAP11B-uniprot.txt]

and the entry carries a CAUTION block (lines 154-159 of the cached record) stating that
although the protein contains a Rho-GAP domain it does not have GTPase activator activity,
citing PubMed:25721503 and PubMed:27957544, and that restoring the activity abolishes the
ability to promote basal-progenitor amplification. That block is paraphrased rather than
quoted because it wraps across CC continuation lines and so cannot be reproduced
verbatim on one line.

The experimental evidence, quoted to the end of the interpreting clause

"However, contrary to ARHGAP11A (27, 28), ARHGAP11B does not exhibit RhoGAP activity when
expressed in monkey fibroblast-like cells (23)." PMID:27957544

(23) is Florio et al. 2015 = PMID:25721503, which is abstract-only in our cache. So the
2016 full text independently confirms what the 2015 IDA records, which matters because I
cannot read the 2015 paper directly.

The important qualification, which must not be dropped:

"Modern ARHGAP11B lacks RhoGAP activity in intact cells (23) and exhibits only very low
RhoGAP activity in a cell-free assay." PMID:27957544

So the honest statement is no activity in cells, very low activity in a cell-free
assay — not "zero under all conditions". NOT|enables is still the right call for an
is_active_in-style claim about the protein in a cell, and it is the curators' call on a
full text I have read, but the review should not overstate it as an absolute.

The decisive genetic experiment is the resurrected ancestral protein:

"In contrast to modern ARHGAP11B, but similar to ARHGAP11A-250, ancestral ARHGAP11B was
found to exhibit RhoGAP activity, as compared to a negative control (empty vector)
(Fig. 2)." PMID:27957544

"However, similar to ARHGAP11A (23), ancestral ARHGAP11B did not exert any effect on
mitotic BP abundance (Fig. 3)." PMID:27957544

i.e. GAP activity and progenitor amplification are anti-correlated across the three
proteins tested. Loss of the catalytic function is not incidental damage; it is what makes
the modern protein do its job.

Why the standard pseudo-enzyme check fails here — and what actually broke

Committed analysis: ARHGAP11B-bioinformatics/ (RESULTS.md, results.json,
regenerated by analyze_arhgap11b.py; --self-test exercises six guards).

The arginine finger is RETAINED. UniProt annotates SITE 87 Arginine finger on
ARHGAP11B and, at the identical position, on ARHGAP11A. The analysis confirms it
reciprocally: both ARHGAP11A's own finger (87) and ARHGAP1/p50RhoGAP's structurally
resolved finger (282) align onto ARHGAP11B residue 87, which is an arginine and is one
of ARHGAP11B's own annotated Site features.

This is worth stating plainly because the campaign's reflex would get it backwards: the
usual "a catalytic term with no catalytic residues is PSEUDOENZYME_OVERANNOTATION"
argument does not apply to this gene, and reaching for it would have produced a
correct action supported by a false reason.

Of 66 reviewed (Swiss-Prot) human proteins carrying the PROSITE RhoGAP profile PS50238,
6 have a non-arginine at their annotated finger position (OCRL, INPP5B/I5P2, ARAP2,
FAM13B with Q; ARHGAP36 with T; DEPDC1B with I).

My first framing of this was "the test works, ARHGAP11B is a false negative of it". Asking
what UniProt independently concluded about those 6 refuted that:

Only one of the six has experimental backing for inactivity. And ARHGAP36 is the
reciprocal of this gene: it lacks the arginine and UniProt calls it an active GAP anyway,
by similarity. So residue identity and curated activity are decoupled in both directions
within this family — which is a stronger and more useful statement than "the test works",
and it is why the residue could not have predicted ARHGAP11B's inactivity. Raised as a
suggested_question against ARHGAP36.

Two bugs on the way to that table, both worth recording:
- I first probed the six using accessions from memory. Q9NYS0 is NKIRAS1, not FAM13B;
the real accession is Q9NYF5, which the census had right because it came from live
UniProt. The measured data was correct and my recollection was not.
- The inactivity test used the bare keyword inactive, which matched ARHGAP36's
"converting them to an inactive GDP-bound state" — the GTPase's state, not the
protein's — and scored a protein UniProt calls active as confirmed inactive. Fixed by
requiring a phrase that can only be about the protein, and regression-tested with that
exact sentence (--self-test T4b).

What the truncation removes instead. Projecting the GAP:GTPase interface from PDB 1TX4
(p50RhoGAP:RhoA:GDP:AlF4, 1.65 Å) onto ARHGAP11A, 2 of 24 mapped interface positions
(ARHGAP11A L221 and A225, both inside its Rho-GAP domain) fall past the divergence point
at residue 220. The catalytic arginine survives; a small part of the surface that has to
present it to the GTPase does not. That is corroboration, not proof — 8% of an interface
is not on its own an argument for inactivity, and the review rests the NOT rows on the
experiments, with the structure as support.

Two measurement bugs were caught by guards and are documented in the analysis README,
because both produced confident wrong numbers first:
- a global alignment put ARHGAP1's arginine finger on ARHGAP11A residue 684 (the two
proteins carry their Rho-GAP domain at opposite ends);
- once fixed, a local alignment reported 0 of 25 contacts lost — because it stopped
at ARHGAP11A residue 204, sixteen residues short of the divergence point. Six contacts
were past its end: not absent, not measured.

Independent reproduction: the divergence point (220) and the prefix identity
(216/220 = 98.2%) are derived from a changepoint fit with no chosen cutoff, and reproduce
the published figures exactly. That is a check on the pipeline, not a new finding.

A UniProt feature defect worth reporting

ARHGAP11B's annotated Rho-GAP domain runs 30 residues past residue 220, i.e. into the
human-specific frameshift C-terminus that is homologous to nothing. The consequence is
perverse: the annotated domain on the inactive truncated paralog is 11 residues longer
than the annotated domain on the catalytically active parent. Any automated step reading
domain extent as evidence of an intact RhoGAP reads this protein as more intact than the
active one. This plausibly contributes to the GO:0007165 IEA from InterPro:IPR000198.

The actual molecular function: mitochondrial, not cytoskeletal

PMID:31883789 is abstract-only in our cache (full_text_available: false), so the IDA
rows sourced from it are accepted on the curators' reading of a full text I have not seen,
which is the correct default. The abstract does state each claim:

"Here, we show that ARHGAP11B is imported into mitochondria, where it interacts with the
adenine nucleotide translocase (ANT) and inhibits the mitochondrial permeability
transition pore (mPTP)." PMID:31883789

"BP expansion by ARHGAP11B requires its presence in mitochondria, and pharmacological
inhibition of ANT function or mPTP opening mimic BP expansion by ARHGAP11B."
PMID:31883789

"Searching for the underlying metabolic basis, we find that BP expansion by ARHGAP11B
requires glutaminolysis, the conversion of glutamine to glutamate for the tricarboxylic
acid (TCA) cycle." PMID:31883789

The two GO:0005515 protein binding IPI rows resolve to UniProtKB:P12235 (SLC25A4 /
ANT1) and UniProtKB:P05141 (SLC25A5 / ANT2) — the ADP/ATP translocases named in that
sentence. These are not screen noise: they are the mechanism, they are corroborated by
UniProt's SUBUNIT line, and they are topologically coherent with a mitochondrial-matrix
protein. Each therefore gets a MODIFY to an informative term rather than the bare
protein binding.

Note the direction of the ANT relationship: ARHGAP11B inhibits mPTP opening via ANT.
GO:0035795 negative regulation of mitochondrial membrane permeability captures exactly
that, and is one of the better-fitted terms in this record.

The neocortex phenotype — attribute it to the system it was measured in

Five of the seven primary papers assay ectopic or transgenic expression of the human
protein in a non-human system. That does not make the annotations wrong (the gene
product assayed is the human protein), but the review should not state mouse or ferret
results as measurements of the human protein in human tissue.

The human-side expression evidence is real and is in UniProt's DEVELOPMENTAL STAGE block,
which records radial-glia-specific expression in human fetal neocortical progenitors:

"Preferentially expressed in apical and basal radial glia"
[file:human/ARHGAP11B/ARHGAP11B-uniprot.txt]

(The same block adds that ARHGAP11B is not detected in cortical neurons or cortical plate.)

Recall gap: the decisive human loss-of-function paper is absent from GOA

The affinage record's trust gates passed, which certifies only that the citations it gave
are real — it says nothing about what it missed. In this case affinage actually supplied
the paper that GOA lacks:

PMID:36098218 is the only study that perturbs ARHGAP11B in a human system, and it
does so in both directions:

"dominant‐negative inhibition of ARHGAP11B's function by ARHGAP11A220 reduces cycling BP
abundance in human cerebral organoids to the chimpanzee level." PMID:36098218

"by subjecting ARHGAP11A plus ARHGAP11B double‐knockout human forebrain organoids to
either ARHGAP11A or ARHGAP11B rescue, we find that ARHGAP11B is essential to maintain the
level of basal (or outer) radial glia (bRG), the BP type of particular relevance for
neocortex expansion." PMID:36098218

Every GO:0021987 row in GOA is a gain-of-function experiment. The necessity evidence
— and the only human-cell perturbation of this gene anywhere — produces no GO annotation
at all. That is a coverage defect, not an over-annotation defect, and it is the single
most useful thing a curator could act on from this review after the Reactome contradiction.

Checks run, including the ones that came back negative

• Row reconciliation: 17 GOA rows, 17 seeded entries, no collapse — a negative result,
  since the fetch-gene stub is known to merge GO:0005515 partner rows and same-term rows
  from different assigners on other genes. Enforced by
  ARHGAP11B-bioinformatics/check_review_counts.py rather than counted by eye, which also
  rejects duplicate YAML keys (PyYAML keeps the last and discards the earlier one silently,
  so data can vanish before any parsed-document check runs).
• Retraction / erratum / expression-of-concern: checked per-PMID against each cited
  article's own PubMed CommentsCorrections/RefType, not by a publication-type search
  (which cannot see Publisher Corrections). Result recorded per reference below.
• IBA: ARHGAP11B has no IBA row at all, and UniProt records
  PAN-GO; Q3KRB8; 0 GO annotations based on evolutionary models. PAINT deliberately gave
  this gene nothing. Given that PANTHER places ARHGAP11B in PTHR15670:SF4, a subfamily
  named for its paralog ("RHO GTPASE-ACTIVATING PROTEIN 11A"), that restraint is the
  correct outcome and deserves to be recorded as such — the predicted IBA-sweep failure
  mode did not occur here. The paralog leak arrived through Reactome instead.
• Arginine-finger loss: hypothesis not confirmed; the residue is retained (above).
• Partner accession resolution: both GO:0005515 partners resolve to reviewed
  canonical Swiss-Prot entries of the expected length, with no TrEMBL/ORFeome substitution.

Final actions

11 + 2 + 4 + 1 = 18 = 17 GOA rows + 1 NEW row.

These counts come from ARHGAP11B-bioinformatics/check_review_counts.py, not from counting
by eye, and writing it immediately earned its keep. My hand tally said ACCEPT 12 and
wrote the arithmetic as 12 + 2 + 4 = 18, which also came to 18 — because it had dropped
the NEW row from the sum at the same time as inflating ACCEPT by one. Two errors cancelling
into a total that looked reconciled. The row list beside the number was correct all along
(2 + 1 + 2 + 1 + 4 + 1 = 11); only the number was wrong. The wrong figure had already
reached the commit message and the PR body before the script was run.

Every REMOVE is non-experimental (three TAS, one IEA). No experimental annotation was
removed, modified or demoted anywhere in this review — the two GO:0005515 MODIFYs
replace an uninformative term with a more specific one and do not weaken any claim.

A tooling limitation that shaped this file: no file: references

This review carries no file: references at all, which is unusual for a gene with a
committed bioinformatics analysis. The reason is environmental, not editorial.

The pre-write hook .claude/hooks/validate_ai_review_pretool_hook.py computes
project_root = Path(__file__).parent.parent.parent and runs the validator with that as
cwd. For an agent working in .claude/worktrees/<id>/, that resolves to the main
checkout, where genes/human/ARHGAP11B/ does not exist — so every file: reference is
rejected as pointing at a non-existent file and the write is blocked.

I measured that this is a false positive rather than assuming it: a copy of this review with
file:human/ARHGAP11B/ARHGAP11B-bioinformatics/RESULTS.md added validates ✓ Valid when
ai-gene-review validate is run from this worktree, and CI runs from a checkout root where
the gene folder exists. So the references would have been fine on the PR; only the local
hook could not see them.

Consequence: the bioinformatics numbers are cited in review.reason prose, naming the
analysis directory, rather than as supported_by entries. Since CI does not check file:
supporting_text at all, nothing verifiable was lost — but the ⚠ No annotations reference available deep research files warning that just validate emits is a direct result of this
and should not be read as the affinage record having been ignored.

2026-10-05 update: the Codex environment validates file: references from the checkout root,
so the review now carries a file:human/ARHGAP11B/ARHGAP11B-bioinformatics/RESULTS.md
reference. The PMID-supported annotation decisions are unchanged; the file reference records
the residue, interface and domain-boundary audit that explains why the propagated RhoGAP
activity and signalling calls fail even though the catalytic arginine is retained.

Open items and limits on what could be verified

• PMID:31883789 and PMID:25721503 are abstract-only. Their IDA rows are accepted on the
  curators' authority; I did not attempt to overrule any of them. Where the abstract is less
  specific than the term it supports — it says "mitochondria", the term says "mitochondrial
  matrix" — the row says so explicitly and defers to the curator's reading of the full text
  rather than letting a quote appear to carry more than it does.
• PMID:36098218's full text is cached locally from PMC, but the reference validator
  resolves it to abstract-only (it tries the publisher copy, which returns HTTP 403). All
  quotes from it are therefore taken from the abstract. This cost a round: the first draft
  quoted the full-text wording and failed validation.
• PMID:32554627 (marmoset, Science 2020) failed to fetch on the first attempt and
  succeeded on a retry; it is now cached (abstract-only) and cited with a verified quote. It
  supports no GOA row, so nothing in the review turns on it. Noted because a transient fetch
  failure is easy to mistake for an unavailable paper.
• GO:0141110 transporter inhibitor activity may be the better molecular function for
  the ANT interaction than the GO:0044325 proposed here, but deciding that needs the full
  text of PMID:31883789 — "inhibits the mPTP" and "inhibits ANT-mediated transport" are
  different claims. Left as a question rather than resolved.