ARHGEF18 / p114RhoGEF — curation notes

Journal for the PAINT-campaign review of human ARHGEF18 (UniProt Q6ZSZ5, HGNC:17090).

Deep-research provider

affinage (human-only), run 2026-09-19 against the 2026-06-09 Affinage record.
Blocking trust gates (accession match, human organism) passed; the soft gate tripped —
Affinage's own head-to-head self-evaluation scored this record pairwise = tie rather
than win against the curated UniProt reference. Recorded in
references[].reference_review for file:human/ARHGEF18/ARHGEF18-deep-research-affinage.md.

A clean gate is a precision claim, and that is how it behaved here: every PMID Affinage
returned resolves to a real paper about this gene. Its recall was the problem.

What the provider missed, and why

missed paper what it decides why a symbol search misses it
PMID:11085924 Blomquist 2000 the founding in-vitro exchange assay; RhoA yes, Rac1 and Cdc42 no title names the function ("a novel Rho-specific guanine nucleotide exchange factor"), never the gene
PMID:15558029 Nagata & Inagaki 2005 the source of UniProt's AltName SA-RhoGEF; SEPT9b binds and inhibits it title names the partner (Sept9b)
PMID:29876405 Chen 2018 1.4 Å structure of the p114RhoGEF PH domain bound to RhoA (PDB 6BCB) title names the family ("Lbc family of RhoGEFs")
PMID:23493395 Medina 2013 activated RhoA binds the PH domains of all seven Lbc RhoGEFs — positive feedback title names the family
PMID:28536193 Schell 2017 EPB41L5 binds and recruits ARHGEF18 in podocytes title names a different protein, and the abstract misspells the gene as "ARGHEF18"
PMID:36912772 Safavian 2023 SEPTIN9 binds and activates ARHGEF18 at the ciliary base title names the partner ("Septin-mediated")
PMID:31409654 Silver 2019 Drosophila ortholog Cysts recruited by Crumbs/Bazooka, activates Rho1 title names the fly gene

Both of the first two are in UniProt's own RN list, which is why reading it is not
optional. The Schell case is the misspelling trap in its purest form: the sentence in
PubMed reads "by binding and recruiting the RhoGEF ARGHEF18 to the leading edge".

Free-supplementary check: PMID:11085924 is free at PMC1221462 and its abstract carries
the decisive substrate sentence, so the paywall on the Circ Res paper (PMID:14512443,
no free route found) does not block the comparison.

The substrate disagreement

Two primary papers, both cited by UniProt, disagree:

A third, independent in-cell result sides with Blomquist: PMID:21258369 — "depletion of
p114RhoGEF resulted in reduced levels of active RhoA in HCE cells without significantly
affecting Rac and Cdc42". Both papers agree on Cdc42 (no).

UniProt carries both, and so should we. The sequence analysis in
ARHGEF18-bioinformatics/ was run to see whether residues could arbitrate; they cannot
(see RESULTS.md §1.3). The one systematic Dbl-family exchange-kinetics panel does not
include p114RhoGEF, so the disagreement has never been retested with modern methods.

What GO can and cannot say

Checked across three independent services (QuickGO, OLS4, the GO API) because QuickGO
silently resolves merges and cannot be read at face value in either direction:

This is the same decision GO made for the RhoGAPs (all merged into GO:0005096). So the
answer to "which GTPase" is inexpressible in the molecular-function branch by design, and
proposing a replacement would be asking GO to undo a deliberate merge. Substrate identity
is carried instead in core_functions[].substrates and the limitation is recorded as an
ONTOLOGY knowledge gap.

The biological-process branch is only slightly better: GO:0035023 retains is_a
children (GO:0035024, GO:0035025, and — asymmetrically — GO:0032489 regulation of
Cdc42 protein signal transduction), but GO:0035025 positive regulation of Rho protein
signal transduction has none, so there is no RhoA-specific or Rac1-specific process term
either. What GO can express and GOA does not is the direction: this protein is a
GEF and every primary paper shows activation, so GO:0035025 rather than the
sign-agnostic GO:0035023. UniProt itself already phrases the RP78 variant effect as
"decreased function in positive regulation of Rho protein signal transduction".

Disease and tissue, kept separate from molecular function

RP78 (MIM:617433, autosomal recessive adult-onset retinal degeneration) is caused by
biallelic ARHGEF18 variants (PMID:28132693). That is a phenotype, and it does not by
itself license any molecular-function or process annotation. The one thing in that paper
that does bear on function is the missense allele: "the p.Thr270Ala missense variant
affects a highly conserved residue in the DBL homology domain, which is required for the
interaction and activation of RHOA". UniProt records it as VARIANT 458 T->A.

The tissue work likewise stays attributed to its system:

None of these is the molecular function. All of them are the same GEF acting in different
places, which is exactly why the location and partner annotations matter more here than
another process term.

The stress-fibre sign flip

GO:0051497 negative regulation of stress fiber assembly (IMP, PMID:25753039) looks
backwards for a RhoA activator, and it is worth being explicit about why it is not an
error. The two observations are made in different systems:

Terry's own explanation: "RhoA activity is downregulated in response to cell confluence;
hence, interfering with junction formation stimulates RhoA signalling in the rest of the
cell." So the negative regulation is a downstream consequence of failing to build
junctions, not a direct molecular action — non-core, kept.

PAINT propagation

Three IBD nodes, from interpro/panther/PTHR47440/PTHR47440-paint.tsv and the GOA
WITH/FROM fields, every token resolved against its own authority:

A rule I got wrong on the first pass

I initially marked the target's own accession CIRCULAR_OR_REDUNDANT in the GO:0005085
and GO:0007264 propagation reviews, and described the GO:0005886 node as a "round
trip". CLAUDE.md prohibits exactly this, and it is right to:

The target appearing in its own WITH/FROM is correct and expected — not circular.
[…] that annotation is one of the descendant evidences the curator used to place the
IBD […] Never mark such a source CIRCULAR_OR_REDUNDANT or describe it as
inflating support.

An IBD is a curator's judgment about where in the tree a function arose, made by reading
the experimental annotations of all extant members — so the target's own IDA being in
the donor list is the marker that experimental grounding exists on the target, and the
IBA then adds the separate claim that the function is inherited rather than
lineage-specific. Nor is a short donor list weak support. What is true, and all that is
true, is that these rows add no evidence the gene did not already have, which makes them
non-core rather than defective. The reviewer caught this; the reason prose had already
been saying the correct thing while the enum said something else, which is the failure
mode to watch for, since nothing in the repo validates agreement between the two.

Which tight-junction term the imaging actually supports

The first draft proposed GO:0005923 bicellular tight junction from Terry's occludin
overlap. That is one level too specific. Confocal colocalisation along a junctional belt
does not separate bicellular from tricellular contacts. GO:0070160 tight junction is the
level the experiment supports, and both QuickGO and OLS4 report it current.

The occludin half of the argument needed correcting once I read the source rather than
citing it. I first wrote that occludin "is not confined to bicellular junctions". What
PMID:40878853 actually shows is narrower: occludin is a BioID proximity partner of
tricellulin at cell-cell junctions and its knockout perturbs tricellulin localisation at
tricellular tight junctions, but "neither one of these other proteins displays similar
prominent accumulation to tTJs as tricellulin" — so occludin is predominantly
bicellular while still being present and functionally required at tricellular contacts.
That is enough to say occludin cannot license the bicellular subtype, and not enough to
say it is indifferent between them. The term choice does not depend on it either way.

The term choice also decides whether the companion GO:0043296 row is informative, which
is worth recording because it is not obvious:

So moving to the parent term is what makes the pair genuinely complementary. Checked in
both services rather than assumed.

Coverage, not over-annotation

ARHGEF18-bioinformatics/reference_coverage.py queries QuickGO by reference,
species-blind: 19 of 23 primary ARHGEF18 papers have produced no GO annotation on any
gene in any species, none undetermined. Everything GOA does assert about this gene is
defensible; the defect is the size of what is absent — the founding biochemistry, the
Nat Cell Biol junction paper, the AJHG disease paper, the only structure of the
protein, and an entire vertebrate knockout. Meanwhile the Drosophila ortholog cyst
is curated (GO:0005085 IDA, GO:0090688 IDA, GO:1902408 IEP from PMID:36917931),
annotations that could seed ISS or IBA to the human gene and have not.

And there is no mammalian ortholog record to lean on. The same script audits the
ortholog records and finds mouse Arhgef18 carries 23 GO annotations of which none is
experimental — ISO 10, IEA 6, IBA 4, ISS 3, neither set truncated — against 14
experimental rows on the human record, which is the positive control for the
evidence-code split. Ten of the mouse rows are ISO from human, so a similarity transfer
back would only recirculate the same four experiments. That matters for PDB 6BCB in
particular: the 1.4 Å RhoA complex was solved on the mouse protein and the mouse
record does not carry it either, so the proposed GO:0031267 has no curated source in
either species and needs creating at the source as well as here.

Residue mapping gotcha

Neither loss-of-function paper states its reference sequence, and they do not agree:

Applying Arno's offset to Terry's position gives an arginine, not a tyrosine, which is
what makes the two assignments distinguishable rather than arbitrary. Tyr-606 is the
pan-Dbl catalytic tyrosine: it is present in all 14 panel members, Rac- and Cdc42-specific
GEFs included, so its retention says nothing about substrate.