ARFGEF3 (BIG3) — computed evidence for the GO review

Two questions, both bearing on GO annotations that human ARFGEF3 currently holds:

1. Is GO:0005085 guanyl-nucleotide exchange factor activity (IBA + IEA) supported
   by ARFGEF3's own Sec7 domain, or is it a fold-derived, name-implied catalytic call?
2. What do the WITH/FROM donors of ARFGEF3's two IBA rows actually contribute?

Everything is fetched live from UniProt/QuickGO and aligned with MAFFT. Nothing is
hardcoded. Re-running reproduces the numbers below; cache/ is disposable.

uv run --no-project --with requests --with biopython python resolve_withfrom.py
uv run --no-project --with requests --with biopython python sec7_catalytic_check.py
uv run --no-project --with requests --with biopython python sec7_catalytic_check.py --self-test

1. The Sec7 glutamic finger is absent from ARFGEF3

Sec7-domain ARF-GEFs activate ARF through one invariant glutamate — the "glutamic
finger" — which inserts into the ARF nucleotide pocket and destabilises bound GDP.
Its presence or absence is therefore the residue-level test of whether a Sec7
domain can do the chemistry its name implies.

Panel construction. The comparison set is not hand-picked: it is the 18
protein donors in the WITH/FROM field of ARFGEF3's own GO:0005085 IBA row, plus
every reviewed ARFGEF3 orthologue, plus a known-dead comparator (below). 17 of the
18 donors have a UniProt SEC7 domain feature and enter the alignment.

Column derivation. The glutamic-finger column is derived, not asserted: it is
the alignment column at which the known-active donors are most Glu-conserved. A
hardcoded column would be a latent bug. The derived column is cross-checked against
an independent, alignment-free signal — the [FY]-x-[LIVM]-P-G-E Sec7 motif — and
the run aborts if the two disagree on any known-active donor.

What the motif cross-check does and does not say. It is corroborating, not
required, and the distinction matters when reading sec7_glutamic_finger.tsv:
12 of the 17 active donors match the motif and place its Glu at exactly the
derived column; 5 match no motif at all (IQSEC2, PSD, Psd3, SYT1, and the
S. pombe protein YDYB_SCHPO); and 0 disagree. Those five are divergent Sec7
branches — BRAG/EFA6 in animals, SYT1 and the fission-yeast protein in fungi — that
retain the catalytic glutamate while departing from the FRLPGE consensus around it,
and every one of them still carries E at the derived column. So a motif_hits: 0
row is silent, not contradictory, and is not counter-evidence to the 17/17
result. The run aborts only on disagreement. These counts are emitted under
motif_crosscheck in sec7_glutamic_finger.json so the point does not rest on
this prose.

Result

Sequence context around the column:

ARFGEF1  (human BIG1,  active)   FRLPGEAQKID      E793
ARFGEF2  (human BIG2,  active)   FRLPGEAQKID      E738
CYTH1    (human,       active)   FRLPGEAQKID      E157
SEC7     (S. cerevisiae, active) FRLPGEGQKID      E923
GNOM     (A. thaliana,  active)  FRLPGESQKIQ      E658
--------------------------------------------------------
ARFGEF3  (human BIG3)            LLSLSNVEEVD      N692   <- no motif at all
Arfgef3  (mouse Big3)            LLALSSVEEVD      S688

ARFGEF3 matches the Sec7 catalytic motif zero times anywhere in its SEC7
domain (583–796). The loss is shared with the mouse orthologue, so it is a
clade-level loss, not a human-specific sequencing or annotation artefact.

What this does and does not license

Losing the catalytic residue is strong evidence against the activity. Retaining
it would not have been evidence for the activity. Only the negative direction is
claimed here. A "retained site" result would have been reported as untested, not
as confirmation — as it was for ARFGEF3's nucleotide-independent regions, which
this analysis does not probe at all.

Controls

--self-test exits 0 with 0 problems, exercising:

• leave-one-out over all 17 active donors — the derived column, and the
  subject's verdict, must survive removal of any single donor, so no one sequence
  is driving the result;
• positive control — a verified-active donor must score "Glu present";
• known-dead control — that same donor with its glutamic finger mutated to Ala
  must score "Glu absent". A detector that cannot report a negative proves nothing,
  so this control also appears in the published table above, not only in the test;
• no-op guard — the mutation target is asserted to be E before it is
  mutated, so a drifted target cannot silently turn the control into a tautology;
• motif guard — the subject must not match the Sec7 catalytic motif.

Reproduction of the published result

Chen et al. 2014 (PMID:24997568) reached the same conclusion from a hand-built
alignment of eight Sec7 domains. Reproducing their panel is a precondition of
this script: it aborts unless BIG3 scores negative and all seven others score
positive. It does — 7/7 and BIG3 = N.

Defect found while reproducing it. That paper's Methods gives GEP100 as
Swiss-Prot:Q6ND90. Q6ND90 is a Rhodopseudomonas palustris succinate
dehydrogenase cytochrome b556 subunit (TrEMBL, no Sec7 domain). Human
GEP100/BRAG2/IQSEC1 is Q6DN90 — two transposed characters. The script records
what each published accession really resolves to and applies the correction
explicitly rather than silently substituting it.

2. The WITH/FROM donors: the IBA node is sound, the target is the exception

withfrom_resolved.tsv resolves all 32 distinct WITH/FROM tokens across the 7 GOA
rows that carry one. Zero unresolved. One (WB:WBGene00007703) resolved only by a
free-text fallback, because WormBase gene ids are absent from UniProt's
xref:wormbase index, which holds protein ids — reported rather than hidden.
Four MOD ids map to more than one UniProt entry and are reported as such, since a
size=1 lookup turns an ambiguity into a confident wrong answer.

ARFGEF3 itself does not appear in either WITH/FROM list. There is no
self-referential donor, so no experimental grounding on the target anywhere in
either propagation chain.

GO:0005085 (guanyl-nucleotide exchange factor activity), 18 protein donors

The three donors without IDA are Drosophila garz (IGI), S. pombe O13690 (IMP),
and S. cerevisiae MON2 (IGI/IPI/ISS).

The node is soundly placed. "This donor only carries the same family-level
inference" is false here in 18 of 18 cases. The defect is entirely on the target
side: ARFGEF3 is the one member of the set that has lost the catalytic residue.
That is root_cause: PROPAGATION_BAD with failure_mode: PSEUDO_OR_SUBACTIVITY_LOSS
— not SOURCE_WEAK_OR_INFERRED, which this analysis directly contradicts.

A second, reciprocal observation worth reporting to PAINT. MON2 is the only
donor UniProt annotates with no SEC7 domain at all, and it is also one of the
three without an IDA. It is a Sec7-family protein in the same structural class as
ARFGEF3 (both carry IPR015403, the Mon2/Sec7/BIG1-like HDS domain). So the same
node appears to be donating GEF activity to a second member that may not perform it.

GO:0016192 (vesicle-mediated transport), 18 protein donors

17/18 carry their own experimental evidence. The donors do not agree on a
specific process: between them they hold GO:0006888, GO:0006890, GO:0006891,
GO:0006892, GO:0006893, GO:0006887, GO:0006895, GO:0042147, GO:0048193,
GO:0048205, GO:0016197, GO:0043001 and GO:0032509 — anterograde ER-to-Golgi,
retrograde, intra-Golgi, endosome-to-Golgi and exocytic steps.

GO:0016192 is therefore the genuine LCA of a heterogeneous donor set, not a
curator failing to be specific. GRANULARITY_MISMATCH requires the donors to
agree; they do not. No specificity upgrade is warranted on this row.

The reciprocal half: the node donates the activity to a second member too

reciprocal_donor_check.py asks what a suspect donor has itself received. Suspects
are chosen by measurement, not by name: any protein donor of GO:0005085 that
lacks the term's own InterPro signature. Exactly one qualifies.

Two of those rows are not independent: the WITH/FROM-set identity is largely
entailed by the shared node, because PAINT emits one descendant-evidence list per
IBD. It is reported as a consistency check on the lookup, not as a second finding.

So one node donates exchange activity to two members that lack a functional Sec7
catalytic site — ARFGEF3, which has the domain but not the glutamate, and MON2, which
has neither — and neither has ever had exchange activity measured directly. MON2's own
support traces to a single paper whose title calls Ysl2p/Mon2 "homologous to Sec7
domain guanine nucleotide exchange factors" and which claims only a "potential
function as an Arf guanine nucleotide exchange factor". That is an architectural and
evidential observation about the node, not about ARFGEF3.

Controls (--self-test, exits 0), each one a separate claim:

• The selector is data-driven. A donor that does carry the signature is never
  chosen as a suspect.
• The count is closed, not a floor. resolve_withfrom.py emits an unresolvable
  MOD id with an empty accession, so a selector that filtered on accession first
  would drop it before any InterPro test — undercounting suspects while still printing
  a confident total. The selector raises on any in-scope token of protein kind that
  failed to resolve, and the self-test clears one accession to prove that guard fires.
  A second tripwire covers a resolved row losing its determination; it is unreachable
  against the current producer and is labelled as such, because an unreachable check
  that reads as coverage is worse than no check.
• The guard does not over-fire. A PANTHER node legitimately has no accession, and
  an unmodified run must stay silent.
• Id normalisation is tested in both directions. MGI:MGI:1334257 and
  MGI:1334257 must compare equal, and two distinct SGD ids must not collapse — a
  normalisation tested only in the merging direction is satisfied by a function that
  returns the empty string.
• The set-identity claim is falsifiable. Removing one member from the subject's
  WITH/FROM set must flip it to False, so a True is a measurement rather than a
  default.
• Committed state is never mutated. Every mutation runs against a copy in a
  temporary directory, and the committed TSV is compared in bytes before and after
  — read_text() normalises newlines while csv.DictWriter emits CRLF, so a text
  comparison cannot see a line-ending rewrite.

3. Two CC rows sit in a disjoint GO branch, because of a SubCell mapping

ARFGEF3's GO:0030133 and GO:0030658 rows come from GO_REF:0000044, the
UniProtKB-SubCell mapping, via SL-0244 and SL-0245. subcell_mapping_check.py
resolves each one and asks how the mapped GO term relates to the term the primary
literature supports — querying QuickGO's is_a/part_of closure rather than
reading it off the labels, because "secretory vesicle" and "transport vesicle"
sound related and are not.

GO:0030141 and GO:0030133 are sibling branches under GO:0031410 cytoplasmic vesicle; neither is an ancestor of the other. So this is not a coarse-but-correct
parent — it is the wrong branch, and the two membrane terms inherit the same split.
GO's definitions make the reason plain: GO:0030133 is "Any of the vesicles of the
constitutive secretory pathway", while insulin and glucagon granules are
regulated secretory granules.

The mismatch starts inside UniProt. SL-0244's own definition is
regulated-pathway language — a vesicle that "mediates the vesicular transport of
cargo - e.g. hormones or neurotransmitters - from an organelle to specific sites at
the cell membrane, where it docks and fuses to release its content". That is
GO:0030141, not GO:0030133.

Scale, stated precisely. SL-0244 is carried by 130,685 UniProtKB entries
and SL-0245 by 90,419 (read from the x-total-results header, not from a page
length). That is the number of entries the mapping is applied to, and it is not
a claim that all of them are mis-annotated: SL-0244 is a broad term and some
proteins in it may genuinely sit in the constitutive pathway. The claim is that the
mapping is wrong for this protein, and that because it is a single vocabulary-level
rule, it would be wrong at that scale wherever the regulated reading applies.

Files