AGFG2 (O95081) — computed analyses supporting the GO annotation review

Numeric tables below are historical outputs of the adjacent scripts/JSON snapshots.
Interpretation was revised on 2026-09-20 after reading the actual PAINT tree and the
GIT biochemical counterexample. Re-run the measurements with the repo virtualenv:

python resolve_withfrom.py     # withfrom.json      — WITH/FROM provenance + donor evidence
python node_reach.py           # node_reach.json    — what each PAINT node donates, to whom
python arfgap_domain.py        # arfgap_domain.json — catalytic-site residues + identity
python provenance_checks.py    # provenance.json    — reference projection, InterPro2GO, family census
python term_checks.py          # term_checks.json   — every term-level fact the review leans on
python distribution.py         # distribution.json  — clade distribution + node taxon breadth
python litsearch.py            # litsearch.json     — recorded PubMed queries + retraction checks
python intact.py               # intact.json        — interaction records, partners, methods
python audit_claims.py         # legacy narrative guard; H requires counts removed from the revised YAML

mafft (v7, --localpair --maxiterate 1000) is the only external binary.

1. Row/annotation reconciliation

AGFG2-goa.tsv has 7 data rows (8 lines including the header), all 7 distinct.
The fetch-gene stub seeded 7 existing_annotations entries. The counts
reconcile exactly; no GO:0005515 or same-term/different-assigner collapse occurred
on this gene, so no rows had to be restored.

2. The catalytic question: two residue observations, unresolved activity

Both the catalytic-arginine position and the ASAP3 Arf-contacting Asp position are
mapped. Neither residue observation alone resolves the activity of AGFG2.

The two positions are located by two different methods, and only one is a
derivation. Both are gated by a literature anchor, but the gates are not equivalent
and the table says which is which:

Only the first row reproduces a literature number from an independent computation. The
second verifies an asserted number and transfers it; the script raises if the residue is
not an Asp, if it falls outside the annotated Arf-GAP domain, or if ASAP3 and AGFG2 fail
to co-align at the catalytic arginine.

Result over the annotated Arf-GAP domains (all Swiss-Prot except drongo):

Every derived arginine falls in the same alignment column, so the comparison is
reciprocally anchored rather than resting on residue identity alone.

The selected panel has 5/5 AGFG proteins lacking Asp and 4/4 non-AGFG ArfGAPs retaining
it. This is a panel result, not a universal biochemical diagnostic. GIT proteins are
not included in the panel: PMID:23433073 reports that all 18 GIT sequences also lack
the homologous Asp, while PMID:10788515 directly measures GIT1/GIT2 ARF GAP activity.

This reproduces, by an independent method and for human AGFG2 specifically, what
PMID:23433073 reports for the subfamily: only two of forty AGFG sequences retain that
aspartate, and the subfamily "is predicted to have lost substantial levels of GAP
activity". That paper's own suggestion is that AGFG proteins are Arf effectors
rather than GAPs.

A refuted assumption, kept on the record. An earlier version of this script placed
the aspartate at a fixed offset (the second of the four residues between the fourth
cysteine and the arginine). Its own control refused that: ASAP3 gave position 466, not
484. The aspartate is 15 residues C-terminal of the arginine in ASAP3, and indels move
it between subfamilies, so alignment transfer is the only sound method. probe_asap3.py
records the measurement that settled it.

What is not shown here. No direct AGFG GAP assay was recovered by the recorded
searches; that is a search/access boundary. Arg75 retention and Thr89 substitution
are verified coordinates, but ASAP3 mutational requirements need not generalize to
every ArfGAP mechanism. The R75Q secretion phenotype is not a GAP assay. The
GO:0005096 inference remains UNDECIDED pending focused biochemical adjudication.

3. Paralogy, orthology, and which one PAINT used

So AGFG1 and AGFG2 are genuine paralogues — same PANTHER family PTHR46134, different
subfamilies, ~48 % identity, both carrying the ArfGAP + FG-repeat architecture — which
is the claim their shared name only implies. PMID:23433073 reaches the same conclusion
phylogenetically: AGFG is one of four subfamilies that "have each undergone a single
duplication resulting in two paralogs".

The closer mouse Agfg2 ortholog is unused as an experimental descendant source. That
is not a negative biological control: PAINT infers function at ancestral nodes from
experimental descendants, then propagates it to descendants without evidence of loss.
Pairwise similarity to the source is not the criterion for accepting the assertion.

4. WITH/FROM provenance and each donor's own evidence

Six distinct WITH/FROM tokens across the 7 GOA rows; zero unresolved. QuickGO
positive control (UniProtKB:P52594 / GO:0001675) returned a non-zero result, so the
zeros below are real zeros and not rejected queries.

Resolver notes reported rather than hidden:

• FB:FBgn0020304 (drongo) resolves to 7 UniProt entries, 0 of them reviewed. Its
  GO:0005737 IDA is therefore real curated evidence attached to a TrEMBL entry; the
  entry's name carries no Swiss-Prot authority.
• MGI:MGI:1333754 resolves to 5 mouse Agfg1 entries, 1 reviewed (Q8K2K6). All
  candidates were queried, not one picked by size=1.
• PANTHER:PTN… tokens identify the actual ancestral assertion. They are relevant
  evidence of phylogenetic transfer, not irrelevant tokens because they are not proteins.

5. Which node reaches which genes, and what it gives them

The recipient counts above describe a frozen query and do not themselves recover
topology. The actual current PTHR46134 tree was subsequently read: target O95081 leaf
PTN002509056 descends from PTN002353603 and PTN002919572. The exact path, current IBDs
and full-response hash are in ../AGFG2-O95081-paint-lineage.json. No target-path loss
was established. All five inherited annotations are retained. Mouse Agfg1
experiments provide descendant grounding, and a short source list is not an objection
to the curator's ancestral placement. The shared AGFG1/AGFG2 assertions are consistent
with inheritance of capacity across their duplication, not evidence of a mistaken copy.

Record counts: AGFG2 has 7 annotations, 1 experimental (the bulk-proteomics HDA);
AGFG1 has 35, 7 experimental.

6. Provenance of the two non-IBA rows

GO:0016020 membrane, HDA, PMID:19946888. Fully paginated, the reference carries
1142 annotations over 1142 distinct gene products, every one GO:0016020, every one
HDA, every one assigned by UniProt. One NK-cell membrane-proteome survey giving 1142
proteins one identical term is a high-throughput experiment. Its abstract explicitly
includes transient membrane association; scale does not make each observation false.
Entity counts are derived as a distinct set of gene-product ids — an annotation count is
not an entity count. Positive control on the same endpoint and call pattern:
GO_REF:0000033 restricted to human GO:0005096 descendants returns 188 annotations over
188 entities. AGFG2's own feature table has no transmembrane segment, no signal peptide
and no lipid-anchor site; peripheral association does not require any of these features.

GO:0005096 GTPase activator activity, IEA, InterPro:IPR001164. AGFG2 matches four
InterPro entries; each one's interpro2go mapping was looked up separately.

Only IPR001164 supplies the electronic activity assertion. Empty mappings on the
other entries do not constitute negative evidence. The broad domain model does not
resolve the biochemical significance of AGFG-specific divergence, and GIT GAP activity
limits generalization from the Asp substitution. Keep the target activity unresolved.

Family census. All 6 reviewed (Swiss-Prot) members of PTHR46134 — human and
mouse AGFG2, human/mouse/rat/bovine AGFG1 — carry GO:0005096 by IEA GO_REF:0000002,
and not one has any experimental evidence for it (0/6). Reported as the Swiss-Prot
subset: 6 entries out of the family's 3 656 proteins, i.e. 0.16 %; the cached
PTHR46134-entries.csv is built from InterPro's reviewed-only endpoint, so this is not a
statement about the family.

7. Term-level facts, verified against QuickGO

• GO:0005096 is not obsolete and GO:0008060 "ARF GTPase activator activity" is one
  of its secondaryIds — i.e. merged, not absent. Its only live child is GO:1902773 via
  capable_of. GO:0005096 is already maximal; no substrate-specific GAP child exists
  to propose or to MODIFY toward.
• GO:0031410 is a verified descendant of GO:0005737, so the two location rows are
  nested rather than independent.
• GO has no term for von Willebrand factor secretion or Weibel-Palade body exocytosis
  (searches for both return only the CC term GO:0033093). But the class is curatable and
  has been curated: GO:0045055 regulated exocytosis has verified is_a children
  including GO:0002576 platelet degranulation, GO:0043299 leukocyte degranulation,
  GO:0016079 synaptic vesicle exocytosis and GO:0060471 cortical granule exocytosis.
  A more specific term is an ontology discussion; the broad term with cargo extension already represents the demonstrated process.
• GO:0046784 viral mRNA export from host cell nucleus was rejected for the Rev-export
  result on two counts: its definition specifies "intronless viral mRNA" whereas the
  Rev/RRE pathway exports intron-containing transcripts, and the historical query returned 0 human
  annotations. The definition mismatch is the substantive issue; an empty query alone
  does not establish biological exclusion.
• GO:0044794 host-mediated activation of viral process carries 58 human annotations
  over 53 entities including IMP, so it is an actively used host-factor term rather than an
  unused corner of the ontology. The declined alternative GO:1903077 carries 37 over 31.
  All of these counts, including the GO:0046784 zero, are recorded in term_checks.json
  with GO:0045055 (369 over 201) as the non-zero control that makes the zero readable.
• GO:0045109's definition is "Control of the spatial distribution of intermediate
  filaments…", i.e. a genuine cytoskeletal-organisation claim, not a by-product term.

8. Recorded search-derived negatives (each a statement about a query)

Positive controls in the same call pattern are shown alongside, so a zero cannot be a
broken query. Full query strings are in litsearch.json.

Retraction / erratum / expression-of-concern check, read from CommentsCorrections on each
cited article's own PubMed record: none flagged for any of the 12 references relied
on (PMID:9303539, 10613896, 11711676, 14724135, 16765935, 19946888, 21284487,
23433073, 25496667, 26701340, 27654348, 34369554). The set checked is the union of
TITLES_NEEDED and CITED_BY_REVIEW in litsearch.py, and litsearch.json records the
count as n_cited_checked: 12 so the number here cannot drift from the sweep.

9. Checks that came back negative, reported as negatives

• Logical-opposite citation cross-product. No pair of AGFG2's 7 terms is a
  positive/negative regulation pair or otherwise logically opposed, so the cross-product
  test has nothing to intersect. Run, empty.
• IntAct (intact.py / intact.json). 10 records, of which 2 are miRNA–mRNA CLASH
  and 8 are protein–protein over 5 distinct partners: AGFG1, TRIM68 and STARD7 by
  anti tag coip in both BioPlex releases (PMID:28514442, PMID:33961781 — one
  platform, not two independent methods), XPO1 by pull-down (PMID:26673895), and
  Yersinia lcrS by two hybrid pooling (PMID:20711500, MI-score 0.37, a host–pathogen
  screen). That leaves 4 human partners with co-IP or pull-down support, which is the
  figure the review quotes. lcrS drops out on method, not on species — the by-name
  non-human filter in the script is a no-op here and is reported as one rather than
  credited with the exclusion. None of the five appears in GOA: AGFG2 has zero
  GO:0005515 rows, so no per-partner verdicts were needed — but the absence is itself
  a possible interaction-curation lead, not a requirement to add generic protein binding.
• Target in WITH/FROM. No token is AGFG2's own accession. This only describes
  which descendants supplied experiments; the positive ancestral placement itself
  supports inheritance by AGFG2. Target self-inclusion would be legitimate, not circular.

10. A discrepancy between two cited papers, settled

PMID:21284487 states AGFG2 proteins are "present in mammals only", from an analysis of
"the first section of the coding mRNAs". Two independent lines say otherwise:

• PMID:23433073, a dedicated phylogenetic study, places the AGFG duplication among the
  subfamilies that duplicated "at the base of vertebrates";
• a UniProt symbol census (totals read from x-total-results, never from a page), using
  distribution.json's own clade names so the sets do not appear to double-count —
  Aves is a subset of Sauropsida, not a separate tally:

The control is non-zero in every clade, so none of the small agfg2 numbers is a
rejected query. The avian asymmetry is real and worth noting — 1 avian agfg2
against 363 avian agfg1 — but it is a symbol count, and the caveat below applies
to it as much as to the positive result.

A symbol census is a name-matching pipeline's output, not an orthologue count, so on its
own it would not settle this; combined with the phylogenetic result it is enough to say the
mammals-only claim is unsupported. Consistently, PTN002919572 reaches lamprey, hagfish
and zebrafish (section 5).