ARHGEF15 (Ephexin-5 / Vsm-RhoGEF) — review notes

UniProt O94989, 841 aa, HGNC:15590, chr17. Dbl-family RhoGEF: one DH domain (417–601) and
an ephexin-type PH domain (CDD cd01221 PH_ephexin; InterPro IPR047270). PANTHER
PTHR12845:SF7 within the ephexin family PTHR12845.

Quotation convention in this file: every quotation is immediately followed by its source in
square brackets, and ARHGEF15-bioinformatics/check_quotes.py verifies each one verbatim
against the cached publication or repository file — including file: sources, which the
repo's own reference validator does not check. An unattributed quotation is reported as a
problem by that checker, not silently ignored.

The short version

The protein is a Dbl-family exchange factor whose exchange machinery is intact and whose
substrate identity is contested between the systems it has been studied in — and GO can
no longer express substrate identity on the molecular-function side in any case. Several
literatures run in parallel and must not be used to argue each other:

The last row is the single most decision-relevant measurement for this gene, it is on the
human protein at full length (Supplementary Table 1 row 56: Human, 841 aa, cDNA
BAA74938.3/AAH36749.1), and it is absent from GOA. Müller et al. 2020 screened all 145 human RhoGEFs/RhoGAPs against RhoA,
Rac1 and Cdc42; Reactome cites this paper for placing ARHGEF15 in both RHOA GEFs activate RHOA and CDC42 GEFs activate CDC42. The paper is paywalled with no PMC record, but its
supplementary tables are free, which is how the row was read
(ARHGEF15-bioinformatics/muller2020_specificity.py). Affinage did not return it, and
neither does a symbol search: its title is
"Systems analysis of RhoGEF and RhoGAP regulatory proteins reveals spatially organized RAC1 signalling from integrin adhesions"
PMID:32203420 — named for RAC1, the one GTPase ARHGEF15 does not act on. It surfaced
only by following Reactome's own citation out of the cached R-HSA-9013159 entry.

One caveat that the analysis itself produced: ITSN1, a textbook Cdc42 GEF, scores negative
for all three GTPases in the same screen. So a − in that table means "not detected here",
not "does not act on". The Rac1 − for ARHGEF15 is nevertheless corroborated independently
by PMID:21029865 and by UniProt's reading of PMID:36929019.

Existing GO record (31 GOA rows)

Every molecular-function row is GO:0005085 guanyl-nucleotide exchange factor activity.
There is no substrate-specific MF row, and there cannot be one — see the ontology
section. Provenance of the MF rows:

• IDA from PMID:12775584, the human VSMC paper.
• IMP from PMID:36929019, the BSVD5 variants.
• IBA from PANTHER:PTN002656129. The PAINT file
  (interpro/panther/PTHR12845/PTHR12845-paint.tsv) shows the underlying IBD placed at
  taxon:2759 (Eukaryota), seeded by MGI:MGI:3045246, O94989 itself, Q12774 ARHGEF5,
  Q5VV41 ARHGEF16 and Q8N5V2 NGEF. MGI:MGI:3045246 was resolved against
  informatics.jax.org, whose record for that accession is Arhgef15, Rho guanine nucleotide
  exchange factor 15 (mouse); it is not quoted here because MGI pages are not part of the
  repository's checkable reference cache. The target appearing in its own WITH/FROM is the
  expected marker that experimental grounding exists on the target itself; it is not
  circular.
• ISS from UniProtKB:Q5FWH6 (mouse Arhgef15), whose own GO:0005085 is IDA from
  PMID:21029865 — a measurement, not a chain of inferences.
• IEA GO_REF:0000120 citing InterPro:IPR000219 (DH domain) and IPR047271
  (Ephexin-like). Weakest of the five, but it points at the right domain: IPR000219 is the
  catalytic DH domain, not an unrelated fold that merely shares a name.

The two ISS rows from UniProtKB:D3ZPJ8 (rat Arhgef15) deserve a note. The rat entry's
GO:0005737 and GO:0051496 are both IDA from PMID:12775584, the same paper that is the
human IDA source. Those rows are not independent corroboration of the human record; they are
the same experiments curated against the rat protein, and the human ISS is a curator
recognising that.

What the literature says, by system

Vascular smooth muscle — the original human identification

PMID:12775584 (Ogita et al. 2003, Circ Res) named the protein Vsm-RhoGEF and reported it as
VSMC-specific:
"a novel VSMC-specific guanine nucleotide exchange factor (GEF) for Rho (Vsm-RhoGEF/KIAA0915) was expressed specifically in VSMCs of several organs including the heart, aorta, liver, kidney, and spleen"
PMID:12775584. The mechanism is receptor-coupled:
"tyrosine phosphorylation of Vsm-RhoGEF induced by EphA4 upon ephrin-A1 stimulation enhanced the Vsm-RhoGEF activity for RhoA"
PMID:12775584, and the downstream requirement rests on two independent perturbations:
"The requirement of Vsm-RhoGEF for ephrin-A1-induced assembly of actin stress fibers in VSMCs was shown by the overexpression of a dominant-negative form of VSM-RhoGEF and by the depletion of Vsm-RhoGEF using RNA interference."
PMID:12775584. The cache is abstract-only; the GOA IDA rests on the full text.

Endothelium — a different substrate, and a contradicted expression claim

PMID:23029280 (Itoh et al. 2012, PLoS ONE) found Arhgef15 as an endothelial-cell-enriched
RhoGEF in mouse retina and reported Cdc42 rather than RhoA:
"In 293T cells, ectopically expressed Arhgef15 as well as its truncated DH-PH proteins induced activation of endogenous Cdc42 (Figure 2A)."
PMID:23029280. One of the two experiments is in human cells:
"VEGF-induced activation of Cdc42 was abrogated by Arhgef15 knockdown in human umbilical vascular ECs (HUVECs, Figure 2B and Figure S4)."
PMID:23029280, so this is not dismissible as a mouse-only observation.

The same paper directly contradicts Ogita's tissue claim, in mouse retina:
"lacZ expression was undetectable in neurons, astrocytes, and vascular smooth muscle cells (vSMCs), ensuring the endothelial specificity of Arhgef15 expression in mouse retinas"
PMID:23029280. Different tissue, different species, different assay — recorded as a real
tension, not as a refutation of either. UniProt still carries Ogita's TISSUE SPECIFICITY line
unqualified:
"Expressed in the vascular smooth muscle of coronary artery."
[file:human/ARHGEF15/ARHGEF15-uniprot.txt].

Neuron — where the substrate claim was made, and where it was later revised

PMID:21029865 (Margolis et al. 2010, Cell) is the paper the mouse ISS and Ensembl rows come
from, and affinage did not return it — its title names EphB and Ephexin5, never
ARHGEF15. It reports the substrate panel, with Ephexin1 as the positive control for Rac1 and
Cdc42 in the same GST-PBD pulldown:
"We conclude that E5 activates RhoA but not Rac1 or Cdc42 (Fig S2A)."
PMID:21029865. Critically, the RhoA assignment was prompted by a sequence criterion, not
discovered by assay:
"Structure-function studies of GEFs identified amino acid residues in the activation domain of Rho family GEFs that specifically identify the GEFs as activators of RhoA rather than Rac or Cdc42."
PMID:21029865. The same criterion produced the GEF-dead control:
"we mutated these three conserved amino acids (L562, Q566, and R567) to alanine (E5-LQR)"
PMID:21029865. The paper also establishes the EphB2 phosphosite, in its own section
heading:
"EphB mediates phosphorylation of Ephexin5 at tyrosine-361"
PMID:21029865.

PMID:40138406 (Petshow et al. 2025, Sci Adv) revises the substrate claim in the same system,
and affinage missed this one too:
"Here, we show that Ephexin5 activates both RhoA and Cdc42 in the brain."
PMID:40138406, with
"The selectivity of Ephexin5 for Cdc42 activation is regulated by tyrosine phosphorylation, which is regulated by neuronal activity."
PMID:40138406. The switch is the residue Margolis identified:
"a mutation of a key tyrosine residue on E5, Y361, enhances E5-dependent activation of Cdc42"
PMID:40138406. They use Margolis's own construct as their GEF-dead control for Cdc42 —
"either E5WT, E5Y361F, or GEF-dead E5 (E5LQR)"
PMID:40138406 — which is what makes the triad exchange machinery rather than a substrate
selector.

PMID:28185854 (Hamilton et al. 2017) grounds the mouse synapse rows that Ensembl Compara
projects into human:
"We found that reducing Ephexin5 levels increased spine outgrowth, and increasing Ephexin5 levels decreased spine outgrowth in a GEF-dependent manner, suggesting that Ephexin5 acts as an inhibitor of spine outgrowth."
PMID:28185854.

Disease — including the one measurement on the human protein outside PMID:12775584

PMID:36929019 (Ding et al. 2023, Acta Neuropathol) established BSVD5:
"In vitro experiments indicated that ARHGEF15 mutations resulted in RhoA/ROCK2 inactivation-induced F-actin cytoskeleton disorganization in vascular smooth muscle cells and endothelial cells and osteoblast dysfunction by inhibiting the Wnt/β-catenin signaling pathway in osteoblast cells."
PMID:36929019. UniProt additionally records from this paper that the protein
"Does not activate RAC1 or CDC42"
[file:human/ARHGEF15/ARHGEF15-uniprot.txt] — a statement in the full text, not the abstract;
the cache is abstract-only, so I do not second-guess it.

PMID:23647072 (Veeramah et al. 2013, Epilepsia) is in UniProt's reference list but absent
from GOA entirely, and it contains a direct assay on the human protein:
"We find that RhoA activation by R604C is reduced by ~46% compared to WT (Figure 3A)."
PMID:23647072. It also states the human-to-mouse residue correspondence that the alignment
here recomputes independently:
"We also find that a homologous amino acid in mouse Ephexin5 at Arg612Cys (R612C) shows a ~47% reduction as compared with mouse WT protein (Figure 3B)."
PMID:23647072. This paper cannot be added as a NEW row, because GO:0005085 is already
in GOA and NEW is rejected for a term already present; it is cited as supporting evidence
on the rows that exist.

Bioinformatics (ARHGEF15-bioinformatics/RESULTS.md)

1. Every published functional residue is retained in the human protein. Mouse
   L562/Q566/R567 map to human L554/Q558/R559 and mouse Y361 to human Y353, which is
   the residue UniProt already annotates as the EphB2 site by similarity. The three
   disease/uncertain variants map cleanly to mouse (R21→R21, V360→V368, R604→R612)
   and all lie outside the DH domain. The V360/V368 mapping resolves the apparent mismatch
   between UniProt's human V360M and the paper's mouse Arhgef15-e(V368M) line, and the
   R604/R612 mapping is confirmed independently by PMID:23647072's own sentence quoted above.
2. The triad does not discriminate substrate in a twelve-protein panel. 0/3
   RhoA-specific GEFs (p115RhoGEF, PDZ-RhoGEF, LARG) carry it; 2/3 Cdc42-specific GEFs
   (FGD1, ITSN1) do. Q and R are near-invariant across all twelve DH domains, so only the
   first column discriminates at all, and there human Ephexin5 groups with the Cdc42-specific
   GEFs. Summary sentence from the results file:
   "The signature points the wrong way in this panel."
   [file:human/ARHGEF15/ARHGEF15-bioinformatics/RESULTS.md]

This is not a refutation of Margolis et al., whose comparison set was structure-based and
different; it is a failure to reproduce the discrimination, which is enough to stop the
residues being used to infer human substrate specificity. Residue identity and measured
substrate are decoupled here in both directions: retention does not establish activity,
and the signature does not establish specificity.

Ontology limitation (cross-checked on QuickGO and OLS4)

GO:0005089 Rho guanyl-nucleotide exchange factor activity no longer exists as a term. It,
together with GO:0005086 (ARF), GO:0005087 (Ran), GO:0005088 (Ras) and GO:0008321
(Ral), has been merged into GO:0005085 — agreed by both services, zero disagreements.
GO:0005085 has zero is_a children; the substrate names survive only as synonyms of the
general term. This mirrors the GAP side exactly, where GO:0005097/0005099/0005100/0008060
merged into GO:0005096 (verified the same way, both services).

Two services were used because neither is reliable alone: QuickGO silently returns the
successor's record for a merged id, so its isObsolete: false is ambiguous between
"current" and "merged away", while OLS4 reports is_obsolete and term_replaced_by against
the id actually requested.

The biological-process branch was not flattened: GO:0007266, GO:0032488, GO:0035023,
GO:0035025, GO:0032489, GO:0035020 and GO:0035022 are all current on both services.
Substrate identity is therefore expressible for this gene only as a BP statement, with two
caveats that matter here:

• GO:0032488 Cdc42 protein signal transduction is an is_a descendant of GO:0007266 Rho protein signal transduction, so GO:0035025 — which ARHGEF15 already carries — is
  satisfied by activating either RhoA or Cdc42 and does not distinguish them.
• GO:0032489 regulation of Cdc42 protein signal transduction has zero children on both
  services: there is no positive regulation of Cdc42 signal transduction term, while Rho
  and Rac both have signed children. A sibling ephexin already sits on the unsigned parent —
  ARHGEF16/Ephexin4 (Q5VV41) carries GO:0032489 by IDA (PMID:21139582) and by IBA from
  PANTHER:PTN002656172, a different PAINT node from ARHGEF15's PTN002656129. PAINT has
  therefore already judged the Cdc42 link to be Ephexin4-specific, and PMID:40138406 is new
  information for that judgement, in mouse.

Provider assessment (affinage)

self_evaluation_pairwise: win, 11 citations, trust gates clear — and the gates are a
precision statement only. The report missed three of the papers this review turns on:

• PMID:21029865 — the Cell paper behind every mouse-derived GO row and the source of the
  RhoA-specificity claim, the triad and the GEF-dead mutant. Titled for EphB and Ephexin5.
• PMID:40138406 — the 2025 Sci Adv paper that revises the substrate claim to RhoA and
  Cdc42.
• PMID:28185854 — the paper grounding the mouse postsynapse, glutamatergic synapse and
  regulation-of-postsynapse-assembly rows that Ensembl projects into human.
• PMID:32203420 — the family-wide specificity screen that measures the substrate range
  directly. Titled for RAC1 and integrin adhesions.

The first three are reachable by searching the alias Ephexin5 rather than the HGNC symbol,
the same recall failure already recorded for ADGB, ADAMTSL1 and ACTG2. The fourth is not
reachable by any gene-name search at all, because the gene is named nowhere in the title,
abstract or PMC-indexed text — only in a supplementary spreadsheet. It was found by
following a citation out of the cached Reactome entry for a GOA TAS row, which is worth
recording as a retrieval route: the TAS rows other reviews treat as low-value provenance
were the only pointer to the best measurement available for this gene.

What affinage did contribute that nothing else surfaced: the endothelial Cdc42/RhoJ
literature (PMID:23029280) and the Sertoli-cell work, neither of which is in GOA. Its
mechanism_profile GO grounding is empty for molecular activity and localization, so nothing
was imported from it.

Decisions taken

• MF. All GO:0005085 rows accepted. The activity is real, measured on the human protein
  in two independent transfected-cell assays (PMID:12775584 IDA; PMID:23647072, not in GOA),
  and the exchange machinery is intact. Substrate identity is carried in
  core_functions[].substrates and in RO:0002233 has_input extensions, because GO's MF
  branch cannot carry it.
• GO:0005515 protein binding. Eight IPI rows. Seven come from binary-interactome
  screens across five distinct partners (PIN1, LASP1, CEP55 x3, PRKG1, GORASP2;
  PMID:25416956, PMID:26871637, PMID:32296183) and are REMOVE. My first pass used
  MARK_AS_OVER_ANNOTATED and argued that a reproducible screen hit is not a refuted
  finding; .claude/skills/annotation-reviewer/SKILL.md:187-199 forbids that action for this
  term and pre-empts that argument in its final clause. The problem with a bare GO:0005515
  is that it carries no functional information, not that it overstates the evidence, so the
  policy routes to MODIFY when the paper supports a better MF and otherwise to REMOVE —
  and removal says the annotation is uninformative, not that the interaction is false. The
  five partners are named in suggested_questions so the removals do not lose them. The
  eighth row, EPHA4 (PMID:12775584), is the MODIFY branch of the same policy:
  GO:0046875 ephrin receptor binding, the receptor-coupling step, a term NGEF/Ephexin1
  already carries for the same role.
• Neuronal rows. Kept as non-core. Every one traces to a mouse experiment
  (PMID:21029865, PMID:28185854) with no human measurement, while the human tissue evidence is
  vascular and adipose. KEEP_AS_NON_CORE is the honest position: the mouse evidence is
  strong and the human relevance is unestablished.
• No NEW rows. Everything the missed literature adds is either already covered by an
  existing term (GO:0005085, GO:0035025, GO:0032956) or inexpressible — there is no
  positive-regulation-of-Cdc42 term. The Cdc42 evidence is mouse brain and mouse retina plus
  one human knockdown; that is a knowledge gap and a suggested_questions item, not an
  assertion.