ARFGEF2 (BIG2) — provenance audit results

All figures below are produced by provenance_audit.py and stored in
provenance_audit.json; audit_claims.py re-checks every number quoted here
against that file, so a stale claim fails rather than lingering.

The GOA set has 66 rows: 26 IEA, 12 IDA, 10 IPI, 7 ISS, 4 IBA, 4 IMP, 2 TAS,
1 HDA — i.e. 27 experimental rows, of which 10 are bare GO:0005515
protein binding. By aspect: 34 cellular_component, 20 molecular_function,
12 biological_process.

1. Thirty percent of the GOA set is a double projection of one rat paper

20 of the 66 rows, covering 13 distinct GO terms, carry rat Arfgef2
(UniProtKB:Q7TSU1, BIG2_RAT, Swiss-Prot, 1791 aa) in their WITH/FROM column.
Only three rat primary references sit behind all 20, and 17 of the 20 rows
rest solely on PMID:15198677 (Charych et al. 2004, J Neurochem).

The same rat annotations arrive twice, by two independent pipelines:

The eleven terms that rest solely on PMID:15198677 are
GO:0005879 axonemal microtubule, GO:0006887 exocytosis, GO:0031410
cytoplasmic vesicle, GO:0032279 asymmetric synapse, GO:0032280 symmetric
synapse, GO:0043197 dendritic spine, GO:0050811 GABA receptor binding,
GO:0098793 presynapse, GO:0098794 postsynapse, GO:0098978 glutamatergic
synapse and GO:0098982 GABA-ergic synapse.

The projection discriminator comes back NEGATIVE here, and that matters.
PMID:15198677 annotates only 2 entities in the whole of GOA — rat Arfgef2
and its partner rat Gabrb3 — so this is genuine per-gene curation, not a
complex-membership projection distributed across subunits. The 23 annotations
are one gene's figures curated at figure granularity (SynGO contributes 12 rows
across four synaptic terms). The defect, such as it is, is single-source
concentration plus double counting by two pipelines, not phantom evidence.

2. GO:0017022 myosin binding reaches this gene through a BIG1 paper

PMID:15644318 (Saeki et al. 2005) is titled "BIG1 is a binding partner
of myosin IXb…". Across GOA it annotates 7 entities with 18 annotations:

• 6 of those annotations are on human ARFGEF1/BIG1 (Q9Y6D6) — GO:0005085,
  GO:0005096, GO:0005515, GO:0017022, GO:0034260.
• 0 are on human ARFGEF2 (Q9Y6D5).
• One is on rat Arfgef2 (Q7TSU1): GO:0017022 myosin binding, IPI with rat
  Myo9b — the only BIG2-family row the paper produced anywhere.

Human ARFGEF2's GO:0017022 row exists only as the Ensembl Compara
projection of that single rat row. Human ARFGEF1 holds the same term directly by
IPI from the same paper.

The term is nevertheless independently true of BIG2, from a different paper:
PMID:23918382 shows reciprocal co-immunoprecipitation of endogenous BIG2 and
non-muscle myosin heavy chain IIA in HeLa cells, and direct binding between
separately in vitro-synthesised BIG2 and NMHC IIA. So the finding is a broken
evidence chain, not a false term — and the same paper records a genuine
paralog difference: BIG1 binds MYPT1 and PP1cδ directly, BIG2 does not.

What this census does not establish. PMID:15644318 has
full_text_available: false in the cache, so these counts show what curators
did, not what the full text contains — the rat GO:0017022 IPI is an
experimental annotation made by someone who read the paper, and nothing here
contradicts it. The defect being reported is on the human side: ARFGEF2's only
support for this term is a non-experimental projection of a single rat row, while
better, directly human evidence (PMID:23918382) exists and is not cited. That
is why the review classifies the row EVIDENCE_CIRCULAR_OR_REDUNDANT with the
rat source left UNRESOLVED, rather than SOURCE_BAD.

3. The reference-scope check on the remaining rows

PMID:19946888 (NK-cell membrane proteome) assigns GO:0016020 membrane to
1142 distinct entities — the row is a proteome sweep, not a BIG2 result. The
three "n/a" rows exceed QuickGO's pagination ceiling; their entity counts are
reported as unavailable rather than guessed from a partial page, while their
per-gene counts stay exact because they come from targeted queries.

4. The real defect is coverage, not over-annotation

Taking the union of the UniProt entry's own RX PubMed list and the affinage
deep-research citation list — a derived set, not a hand-picked one — gives
43 literature PMIDs. Of those:

• 26 produce zero GO annotations anywhere in GOA;
• 33 produce zero annotations on ARFGEF2.

The 26 silent papers include the gene's founding disease paper
(PMID:14647276, Nat Genet 2004), the myosin-phosphatase scaffold paper
(PMID:23918382, PNAS 2013), integrin β1 recycling and cell migration
(PMID:22908276, PNAS 2012), Filamin A transport (PMID:16320251), the
AP-1/GGA dominant-negative papers (PMID:11777925, PMID:12051703), PP1γ
regulation (PMID:17360629), the RIIβ/TNFR1 AKAP paper (PMID:18625701),
BIG1/BIG2 redundancy in retrograde transport (PMID:18417613), the DCB/HUS
homodimer paper (PMID:17640864), β-catenin S675 (PMID:27162341) and
dendritic Golgi deployment (PMID:29455446).

For a disease gene the campaign's standing expectation is over-annotation from
the disease phenotype. ARFGEF2 inverts it: GOA carries no cerebral-cortex,
neuron-migration, neural-progenitor or heterotopia term at all, and the
pathology-derived over-annotation simply is not there. The predicted failure
mode was looked for and not found; what is there instead is a coverage gap.

5. The IBA node is heterogeneous, so the general GEF term is the LCA

GO:0005085 IBA draws on 19 WITH/FROM tokens — the ancestral node
PANTHER:PTN008950430 plus 18 proteins spanning the whole Sec7 superfamily:
Arabidopsis GNOM, Drosophila Sec71 and
garz, mouse Cytohesin-1 and PSD3, yeast GEA1/GEA2/SEC7/MON2/SYT1, S. pombe
SPAC11E3.11c, and human PSD, CYTH1, CYTH3, CYTH4, IQSEC2, ARFGEF1 — plus
ARFGEF2 itself (UniProtKB:Q9Y6D5), which is the expected marker that the target
carries its own experimental grounding for the term, not a circularity.

Because that clade mixes BIG-type, GBF-type, cytohesin-type and MON2-type GEFs
acting on different ARF/ARF-like substrates, guanyl-nucleotide exchange factor activity is the least common ancestor of the donor set, not a curator
failing to be specific.

Independently: GO:0005086 ARF guanyl-nucleotide exchange factor activity no
longer exists as a distinct term. QuickGO's complete record for GO:0005085
lists GO:0005086-GO:0005090, GO:0008321, GO:0008433, GO:0016219,
GO:0016220, GO:0017034, GO:0017112, GO:0017132, GO:0019839 and
GO:0030676 among its secondaryIds — every substrate-specific GEF term was
merged into it. So GO:0005085 is already maximal for this protein and there
is no child to propose; the ARF1/ARF3 substrate preference has to be recorded
machine-readably (core_functions[].substrates, and a has_input extension on
the annotation) rather than as a more specific term.

6. Cilium census: is a ciliary term plausible anywhere near this protein?

The two GO:0005879 axonemal microtubule rows claim a cilium. Rather than assert
"no other ArfGEF has one", the claim was measured. The cohort is the four human
large ArfGEFs — accessions derived by gene-name lookup, not written by hand,
after a first pass hardcoded Q9Y678 as GBF1 when it is in fact COPG1 — plus
the nine proteins resolvable from this gene's own WITH/FROM column. Thirteen
accessions; each was queried for GO:0005879 and the cilium/axoneme compartment
terms an ArfGEF would plausibly receive if it were ciliary.

Two results, and the second is the one that complicates the story:

1. Of the four human large ArfGEFs, only ARFGEF2 holds a cilium-compartment
   term, and only by the projection under review. The family gives no
   independent support to a ciliary localisation.
2. EXOC7/Exo70 — the exocyst subunit BIG2 binds and co-localises with at the
   MTOC — carries GO:0036064 ciliary basal body. So a ciliary context near
   BIG2 is not absurd, and the honest verdict is "flag and measure", not "delete".

7. WITH/FROM resolution

All 45 distinct WITH/FROM tokens across the 47 rows that carry one resolved;
0 unresolved. Five MOD ids returned more than one UniProt hit
(RGD:1560793, RGD:631430, FB:FBgn0264560, MGI:MGI:1334257,
MGI:MGI:1918215) — in every case all hits are the same gene (the Swiss-Prot
canonical entry plus TrEMBL isoform records), so identity is unambiguous; the
multiplicity is reported rather than collapsed. Two tokens resolve only to
unreviewed (TrEMBL) entries: FB:FBgn0028538 (Drosophila Sec71) and
FB:FBgn0264560 (Drosophila garz). Their annotations may still be
experimental, but their protein names are automatic, so neither name is used
here as evidence of what the family does.