UniProt Q5TH69, 2177 aa, HGNC:21213. Synonyms BIG3, C6orf92, KIAA1244.
Pharos class Tdark; PE 1: Evidence at protein level (detected by MS in
several proteomes, and by western/IHC in breast cancer tissue).
The gene symbol, the UniProt recommended name ("Brefeldin A-inhibited guanine
nucleotide-exchange protein 3", itself ECO:0000305 — a curator inference from
similarity, not a measurement) and both GO:0005085 rows all assert ARF-GEF
activity. Nothing has ever measured it on this protein, and the residue that
performs the chemistry is absent.
Sec7-domain ARF-GEFs act through one invariant glutamate, the "glutamic finger",
which inserts into the ARF nucleotide pocket to destabilise bound GDP. Chen et
al. showed BIG3 lacks it:
and the divergence is not marginal:
The group that named the protein says the same thing in their own papers:
ARFGEF3-bioinformatics/I did not take this on the papers' word. Re-derived from live UniProt + MAFFT,
with the comparison panel built from the GOA WITH/FROM field rather than by
hand (sec7_catalytic_check.py, full write-up in
ARFGEF3-bioinformatics/RESULTS.md):
| class | n | residue at the derived glutamic-finger column |
|---|---|---|
| catalytically-verified WITH/FROM donors | 17 | E in 17/17 |
known-dead comparator (ARFGEF1 E793A) |
1 | A — scored absent |
| human ARFGEF3 | 1 | N692 (LLSLSNVEEVD, zero motif matches) |
| mouse Arfgef3 | 1 | S688 (LLALSSVEEVD) |
The column is derived from the known-active donors, not hardcoded, and
cross-checked against the alignment-free [FY]-x-[LIVM]-P-G-E motif. Controls run
in both directions and the --self-test exits 0: leave-one-out over all 17 donors,
a verified-active positive control, and a Glu→Ala mutant of that control which must
score negative (a detector that cannot report a negative proves nothing). The
mutation target is asserted present before it is mutated, so the control cannot
silently no-op.
The asymmetry matters and I have kept to it. Losing the catalytic residue is
strong evidence against the activity; retaining it would not have been evidence
for it. Only the negative direction is claimed.
Reproduction defect found. Chen et al.'s Methods give GEP100 as
Swiss-Prot:Q6ND90, which is a Rhodopseudomonas palustris succinate dehydrogenase
cytochrome b556 subunit (TrEMBL, no Sec7 domain). Human GEP100/BRAG2/IQSEC1 is
Q6DN90 — two transposed characters. The script reports what each published
accession really resolves to and applies the correction explicitly.
Chen et al. write that their prediction is
and the <xref rid> in the PMC XML resolves that citation to PMID:24711543
(Li et al. 2014). I could not locate a GST-GAT pull-down in the accessible text
of that paper — it is neither in the cached full text nor in any of its five named
supplementary figures, and PMC restricts its XML. So the direct experimental
demonstration that BIG3 fails to load ARF with GTP is asserted in the literature
but I could not verify it. The review therefore rests on the residue analysis and
the authors' repeated sequence-level statements, not on that pull-down. Recorded
as a question for the authors rather than quoted as evidence.
Two well-supported activities, neither of them catalytic, and neither in GOA.
The founding human paper:
The mechanism is occlusion of PHB2's import receptors, not a generic tether:
This is GO:0140311 protein sequestering activity almost verbatim — GO defines it
as "Binding to a protein to prevent it from interacting with other partners or to
inhibit its localization to the area of the cell or complex where it is active."
Both clauses are separately demonstrated here: the localisation clause by knockdown
and re-expression, the partner-occlusion clause by the karyopherin work.
Parent vs child, weighed rather than defaulted. GO:0140311 is an is_a child
of GO:0140313 molecular sequestering activity, carrying the identical definition
with "a specific molecule" narrowed to "a protein". PHB2 is a protein, so the child
adds no claim that the ligand's identity does not already fix, and the child is
correct. This is the opposite situation to the PKA row, where GO:0034237 protein
kinase A regulatory subunit binding would add an unproven claim — the
interaction was measured but the RII-subunit specificity was only predicted in
silico — so that row deliberately stays at the parent GO:0051018. The test is not
"prefer the child" or "prefer the parent" but whether the child asserts anything
the evidence has not already fixed.
The same test applies to the phosphatase row: GO:0004865 protein serine/threonine
phosphatase inhibitor activity over its parent GO:0004864, because PPP1CA is
EC 3.1.3.16, a serine/threonine phosphatase by definition.
The downstream process is GO:0042308 negative regulation of protein import into
nucleus.
Interface mapped to a single predicted helix and confirmed by mutagenesis:
Q165/D169/Q173, within the experimentally-delimited PHB2-binding region 86–434
(PMID:24997568; the peptide built from it is ERAP, PMID:24051437). Note that
BIG3 acting on ERα directly was tested and failed:
so ERα-pathway terms would be a ROLE_CONFLATION: BIG3's substrate is PHB2, and ERα
activity is two steps downstream.
Binding is via a canonical PP1C docking motif (RVxF, 1,228-KAVSF-1,232) and the
ΔPP1Cα deletion abolishes it. E2-driven PKA phosphorylation of BIG3-S305/S1208
then relieves the inhibition, so PP1Cα dephosphorylates PHB2-S39. So BIG3 is a
genuine GO:0008157 protein phosphatase 1 binding + GO:0004865 protein serine/threonine phosphatase
inhibitor activity + GO:0051018 protein kinase A binding protein. All three are
human, endogenous-protein experiments, and none is in GOA.
This is also where the family connection survives: BIG1 and BIG2 carry AKAP
sequences too (PMID:28555617 intro). So what BIG3 inherited from the family is the
scaffolding, not the catalysis.
The strongest functional data are from mouse and must be encoded as ISS/ISO, not
IMP/IDA (the experiments were not done in human).
Same sign in alpha cells:
The authors' own model is explicitly that the dead Sec7 domain is the point:
That is a hypothesis, not a result, and I have not annotated it. But it does mean
the right reading of ARFGEF3 is not "a GEF that GO over-called" — it is "a
catalytically dead family member that regulates the pathway its active relatives
catalyse".
UniProt already carries the granule localisation as ECO:0000250|UniProtKB:Q3UGY8
(from mouse), which is where the GO:0030133 / GO:0030658 IEA rows come from.
Those SubCell terms say transport vesicle; the mouse data say secretory granule
and trans-Golgi network, which are different things — a genuine, fixable
mismatch (see the review's GO:0030133 and GO:0030658 rows).
| compartment | cell type | evidence |
|---|---|---|
| cytoplasm (diffuse) | human breast cancer | IHC/ICC, endogenous, PMID:19496786 |
| TGN / immature secretory granules | mouse islet β- and α-cells | PMID:24711543, PMID:25737957 |
| lysosome | mouse hippocampal neurons | PMID:27696409 |
| mitochondrion | human osteosarcoma lines | PMID:34363714 |
The osteosarcoma paper explicitly contrasts its own result with the breast-cancer
localisation, so this is not a contradiction being papered over — it is a
cell-type-dependent distribution. GO:0005737 cytoplasm (the one EXP row) is the
only one of these four that is human and in GOA; the other three are coverage gaps
and are all non-human or non-canonical enough that I have kept them out of
core_functions and raised them as questions instead.
ARFGEF3-deep-research-affinage.md, faith_pct: 100.0, 14 citations, trust gates
clear. The gates are a precision gate; there is no recall gate, and both
failure directions showed up here.
Precision failure the gates do not cover. Two of its 14 citations —
PMID:14657013 and PMID:15707593 — are not about this gene at all. They concern
"BIG-3 = BMP-2-induced gene 3 kb", a 34-kDa, seven-WD-40-repeat protein
(PMID:11551928), i.e. a pure name collision with ARFGEF3/BIG3 (2177 aa, 240 kDa,
Sec7 + ARM, no WD40). The affinage narrative presents their chondrocyte- and
osteoblast-differentiation results as ARFGEF3 findings. Both are marked
relevance: NONE, correctness: MISCITED in the review, and nothing rests on them.
Note the gates still passed: the PMIDs are real and the quotes faithful — the join
to this gene is what is wrong, which is precisely the class of error no mechanical
gate sees.
Recall failures. An independent PubMed sweep over ARFGEF3 OR BIG3 OR BIG-3 OR
KIAA1244 (129 hits, mostly "Brain Injury Guidelines" noise) found four real
ARFGEF3 papers the provider never returned:
So: 2/14 provider citations were the wrong gene, and the single most
compartment-informative paper was absent. Consistent with the campaign measurement
that the gates certify only what was returned.
ARFGEF3-bioinformatics/)All 32 distinct WITH/FROM tokens across the 7 rows that carry one resolve; zero
unresolved. WB:WBGene00007703 resolved only by free-text fallback (WormBase gene
ids are absent from UniProt's xref:wormbase index, which holds protein ids) and is
reported as such. Four MOD ids map to >1 UniProt entry and are reported with their
hit counts rather than collapsed.
ARFGEF3 does not appear in either of its own WITH/FROM lists, so there is no
self-referential donor and no experimental grounding on the target anywhere in
either chain.
GO:0005085 — the node is sound, the target is the exception18 protein donors. 18/18 carry their own experimental evidence for the term, and
15/18 carry their own IDA (direct exchange assay). The tempting dismissal
("these donors only carry the same family-level inference") is factually false here,
in every single case. So SOURCE_WEAK_OR_INFERRED would be contradicted by my own
analysis; the correct classification is PROPAGATION_BAD +
PSEUDO_OR_SUBACTIVITY_LOSS: PAINT placed the activity correctly at an ancestral
node, and ARFGEF3 is the descendant that lost it.
Reciprocal observation worth sending to PAINT, checked rather than conjectured
(reciprocal_donor_check.py): MON2 (SGD:S000005241) is the only donor with
no annotated SEC7 domain and no IPR000904, and is also one of the three
without an IDA (IGI/IPI/ISS only). MON2 and ARFGEF3 share IPR015403
(Mon2/Sec7/BIG1-like HDS).
The script selects suspects by measurement — a protein donor of the term lacking
the term's own InterPro signature — so MON2 is its output, not its input, and
exactly one donor qualifies. MON2 receives GO:0005085 by IBA from
PANTHER:PTN008950430, the same node ARFGEF3's row comes from, and its entire
non-IBA support is one reference, PMID:12052896, with no IDA. That paper is itself
careful — its title calls Ysl2p/Mon2 "homologous to Sec7 domain guanine nucleotide
exchange factors" and it claims only a "potential function as an Arf guanine
nucleotide exchange factor".
So the same node is donating GEF activity to a second member that has never
been shown to perform it — the mis-placed-member pattern, with the donor and the
victim being the same kind of protein. (The WITH/FROM-set identity the script also
reports is largely entailed by the shared node, since PAINT emits one
descendant-evidence list per IBD; it is a consistency check on the lookup rather
than a second independent finding.)
GO:0016192 — the broad term is the correct LCA17/18 donors carry their own experimental evidence, but they do not agree on a
process: between them they hold GO:0006887, GO:0006888, GO:0006890,
GO:0006891, GO:0006892, GO:0006893, GO:0006895, GO:0016197, GO:0032509,
GO:0042147, GO:0043001, GO:0048193 and GO:0048205 — anterograde, retrograde,
intra-Golgi, endosomal and exocytic. GO:0016192 is the genuine LCA of a
heterogeneous donor set. GRANULARITY_MISMATCH requires the donors to agree; they
do not, so no specificity upgrade is proposed on that row and it is ACCEPTed.
Found while checking the two GO_REF:0000044 vesicle rows rather than accepting
them. subcell_mapping_check.py resolves each SubCell id in the WITH/FROM column
and asks how the mapped GO term relates to the term the literature supports —
querying QuickGO's is_a/part_of closure, not reading it off the labels.
| SubCell | UniProt's name | maps to | literature supports | relation |
|---|---|---|---|---|
SL-0086 |
Cytoplasm | GO:0005737 |
— | EXACT |
SL-0244 |
Secretory vesicle | GO:0030133 transport vesicle |
GO:0030141 secretory granule |
DISJOINT |
SL-0245 |
Secretory vesicle membrane | GO:0030658 |
GO:0030667 |
DISJOINT |
GO:0030141 and GO:0030133 are sibling branches under GO:0031410; neither is
an ancestor of the other. GO defines GO:0030133 as the constitutive secretory
pathway, while insulin and glucagon granules are regulated secretory granules.
And the mismatch starts inside UniProt: SL-0244's own definition is
regulated-pathway language ("cargo - e.g. hormones or neurotransmitters ... docks
and fuses to release its content").
So the two rows needed MODIFY (a replacement) rather than a specificity
refinement. SL-0244 is applied to 130,685 UniProtKB entries and SL-0245 to
90,419 (read from the x-total-results header). That is the number the rule is
applied to, not a claim that all of them are mis-annotated — SL-0244 is broad
and some members may genuinely be constitutive.
Recording these because both initially reported success while being blind.
My quote checker was structurally blind to 27 of 65 quotes. It collected
only dicts carrying both reference_id and supporting_text, so every
references[].findings[] quote — where the reference is implicit in the
position — was skipped, and it reported "0 problems" over the 38 it could see.
The raw-vs-parsed reconciliation (65 raw keys vs 38 parsed) is what exposed it.
I derived the expected number independently (27 findings + 32 annotation
supported_by + 6 core_functions = 65) rather than finding a story that made
the gap acceptable, which confirmed the document was intact and the checker
was at fault. Fixed, then mutation-tested on a findings-level quote specifically.
The file: quote checker's mutation test passed by crashing. Exit 1 came
from a RuntimeError in root derivation, not from detection. A check that kills
the harness is worse than none, because it still prints as though it ran. Fixed
with a CWD fallback and re-tested: the mutant now produces a PROBLEM line and
the clean file exits 0.
All five file: quotes were then verified as literal substrings of RESULTS.md.
CI does not check these (file is in skip_prefixes), so they are the one
fabrication surface here, and each is a computed table row chosen so an exact match
is meaningful.
Worth knowing for anyone running agents in isolated worktrees, because it produces
confident false failures.
.claude/hooks/validate_ai_review_pretool_hook.py computes
project_root = Path(__file__).parent.parent.parent and runs the validator with
cwd=project_root. __file__ is the shared checkout's hook, so validation uses
the shared checkout's schema and its genes/ tree — not the worktree the edit is
actually being made in. Two consequences hit this gene:
paint/ABHD8 (tip 0e4e91ae55), which predates4697a4e0c7 "Add FamilyReview schema with machine-checkable residue claims"residue_claims as an unknown property even thoughPropagationReview on main for some time.reference_base_dir: genes resolved against the shared checkout, wheregenes/human/ARFGEF3/ARFGEF3-bioinformatics/RESULTS.md does not exist, so allfile: references were reported as pointing at a non-existent file.Diagnostic that settles it: run the hook's own procedure on a file you know is
good. Copying the unmodified, committed review to a temp directory and running
uv run ai-gene-review validate --no-goa on it returns ✓ Valid. So does
just validate human ARFGEF3 in the worktree. Only the hook disagrees, and the
branch-ancestry check (git merge-base --is-ancestor 4697a4e0c7 0e4e91ae55 → 1)
names the reason.
Same shape as the shared-checkquotes.py-with-a-hardcoded-root incident: when a
shared tool suddenly reports failures on work you believe is clean, suspect the
tool and check what root it resolved before changing anything. "Fixing" these
would have meant deleting two correct residue_claims blocks and five correct
citations. CI is unaffected — it validates against main's schema on real paths.
The first write attempt was blocked with a long error list including six
"Text part not found as substring" failures and several missing-file:-reference
errors. Those were not real. The genuine errors were four type errors
(experts must be strings, substrates must be Term objects,
reference_section_type is not a Reference-level slot, gap_kind is
multivalued); because the generated JSON Schema matches the root with anyOf,
one type failure makes the validator report errors from a mismatched alternative.
Fixing only the four type errors turned the file ✓ Valid with every quote
untouched. Had I "corrected" the six quotes, I would have corrupted correct work —
the shared-tool lesson in a new guise.
paint/ARFGEF1 and paint/ARFGEF2 branches exist but are still at the branch point
with nothing pushed, so I could not compare resolutions. Flagging the divergence in
advance, because all three genes carry the same GO:0005085 IBA from the same
node and it should be resolved differently:
| gene | glutamic finger | expected verdict |
|---|---|---|
| ARFGEF1 (BIG1, Q9Y6D6) | E793 present | ACCEPT — also has its own IDA |
| ARFGEF2 (BIG2, Q9Y6D5) | E738 present | ACCEPT — also has its own IDA |
| ARFGEF3 (BIG3, Q5TH69) | N692 — absent | MARK_AS_OVER_ANNOTATED |
Both paralogue positions were verified against the live UniProt sequences, so this
is a checkable claim rather than an assumption about how those reviews will land.
Checked each PMID relied on against its PubMed record. No retraction, erratum,
expression of concern or publisher correction found on any of them. Recording the
negative so the next reviewer knows the check ran.
GO:0005794 Golgi apparatus row. UniProt's DR GO block lists itIEA:UniProtKB-ARBA) but it is not in the GOA TSV, so it is not an existingGO:0005739 mitochondrion / GO:0005764 lysosome in core_functions. EachGO:0140311, GO:0042308, GO:0008157, GO:0004865, GO:0051018,GO:1904410. All eight candidate ids were checked against QuickGO/complete for isObsolete and secondaryIds before use; none is obsolete orGOA TSV: 8 data rows. fetch-gene stub: 8 existing_annotations entries. They
match — no collapsed GO:0005515 partner rows and no collapsed same-term rows on
this gene, so the under-seeding defect seen elsewhere in this campaign is absent
here. Checked explicitly rather than assumed.
Final review: 14 entries = 8 GOA rows + 6 NEW proposals. Verified
programmatically, not by eye, by arfgef3_check_entities.py, which also asserts
that every row's supporting_entities reproduces the GOA WITH/FROM field exactly
and that every propagation_review names only sources drawn from it. Hand-built
source lists have drifted on every gene in this campaign that tried it.
just validate reports ✓ Valid (with 1 warnings): "No annotations reference
available deep research files". Satisfying it would mean putting
file:...ARFGEF3-deep-research-affinage.md into a supported_by — i.e. quoting a
provider sentence as evidence, which is the practice this campaign specifically
forbids, and on a record that turned out to cite two papers about a different gene.
The provider's contribution is assessed at length above instead. Left standing
rather than silenced.