AGFG1 (P52594) — computational analyses supporting the GO review

Every number below is produced by a script in this directory. Re-run order does not
matter; each script is standalone and fetches its own inputs.

1. The arginine finger and zinc finger are intact — but that is ONE of three required residues

RETRACTION, and it is the headline result of this review. This section originally
concluded "the pseudoenzyme hypothesis is NOT confirmed" on the strength of the
arginine finger being present. That conclusion was an artefact of testing only the
residue the 2008 consensus-nomenclature paper happens to name. The field identifies
three catalytically required positions, AGFG1 retains one, and section 1b measures
it. The verdict on GO:0005096 moved from KEEP_AS_NON_CORE to
MARK_AS_OVER_ANNOTATED as a result. The table below is still correct about what it
tests; it is simply not sufficient, and the sentence above is why. Read 1b before drawing
anything from 1.

The anchor is published, not remembered. PMID:18809720 (the consensus-nomenclature
paper for the 31 human ArfGAPs) states the domain contains "a characteristic C4-type
zinc finger motif and a conserved arginine that is required for activity, within a
particular spacing (CX2CX16CX2CX4R)", and names ArfGAP1's residues as Cys22/25/42/45 +
Arg50.

arfgap_motif.py reproduces those published numbers exactly as a positive control
before reporting anything else, and raises if it cannot.

12/12 panel members retain CX2CX16CX2CX4R.

The last row is the built-in negative control and it is the point of the table: the same
reference states that "Arf GAP activity has been demonstrated in vitro for at least one
member of each subfamily, with the exception of the ADAPs, which appear to lack in vitro
GAP activity" — yet ADAP1's motif is complete. Sequence retention of the apparatus is
necessary and demonstrably not sufficient, so it cannot on its own establish
GO:0005096 for AGFG1, and equally it removes any structural ground for withdrawing it.

Both clauses of that sentence matter, and the first one cuts the other way. Raised in
review, and correctly: the clause before the ADAP exception is a blanket positive that
covers this subfamily, because the paper classifies the 31 human ArfGAPs into ten
subfamilies, AGFG is one of them, and only the ADAPs are excepted. Quoting the ADAP half as
a control while passing over that would be a verbatim-but-selectively-bounded quotation.
Four checkable grounds for discounting its AGFG instance rather than treating it as the
measurement:

1. the clause carries no citation, while the sentences on either side of it cite
   specific papers (Liu et al. 2005; Bowzard et al. 2007);
2. it is stated at subfamily level and names no species, protein or assay;
3. the paper's own AGFG section — where a specific claim would live — says nothing
   about GAP activity for either paralogue, only "Much less information is available on
   AGFG2.";
4. the same paragraph warns that "some ArfGAPs use their GAP domain to bind Arf without
   promoting GTP hydrolysis", which is exactly the distinction a subfamily-level summary
   cannot settle.

So the blanket clause raises the prior that some AGFG member is catalytically active,
without supplying the measurement — and section 1b shows why that prior does not survive
for this subfamily.

1b. Two of the three catalytically required residues are substituted in every AGFG protein tested

catalytic_residues.py. The decisive source is PMID:23433073, which was surfaced by the
concurrent AGFG2 review and which this review verified independently rather than inheriting.
It names three positions in ASAP3 — the ArfGAP with a solved Arf6 complex — and states:

"The three that are found in our analyses (W451, R469, and D484 in ASAP3 correspond to
W14, R32, and D47 in Figure S4) are each closely involved in catalysis. R469/R32 is the
arginine finger. D484/D47 contacts the main chain of Arf6-Q67 plus D68 and the side chain
of Q67, stabilizing switch 2 and catalytic glutamine in Arf6. W451/W14 is centrally
located in the binding interface between the Arf and Arf GAP. Mutation of any one of
these three residues leads to severe loss in Arf GAP activity"

Three controls, all of which must pass before any absence is reported, and the script raises
if they do not:

1. ASAP3's own residues 451/469/484 are still W/R/D in the current UniProt sequence, so
   the paper's numbering has not drifted;
2. every GAP-competent member of the panel must recover all three — if the alignment
   cannot find these residues in proteins known to have them, an absence in AGFG1 would be
   an artefact rather than a result;
3. the arginine this alignment finds must equal the one the independent CX2CX16CX2CX4R
   scan finds — two methods, one residue. 8 of 9 entries agree.

5/5 GAP-competent controls at 3/3; 0/4 AGFG-family members at 3/3; 4/4 retain the
arginine finger alone. This is exactly the subfamily-wide loss the source paper predicts
from 40 AGFG sequences — "Only two of the 40 AGFG sequences contain an aspartate at the
position homologous to D47 in the other subfamilies" and "The AGFG consensus also uniquely
lacks W14" — and its conclusion names the subfamily: "the ArfGAP is a very highly
conserved structural domain that is predicted to have lost substantial levels of GAP
activity in at least one subfamily (AGFG)".

The reversal resolves a tension rather than creating one. Section 14's ARF-proximity
datum (AGFG1 with ARF1, ARF3 and ARF6) had looked like support for catalysis. The same paper
is explicit that it is not: "These predicted changes (including complete loss, potentially)
in GAP activity or its regulation should not be confused with consequent changes in the
ability to bind Arf family GTPases." Binding retained plus catalysis lost is one coherent
picture — an Arf effector, not an Arf GAP.

Cross-review agreement, reached independently. The concurrent paint/AGFG2 review
derived Thr89 for AGFG2 from the same paper; this panel reproduces that number exactly, and
adds AGFG1, mouse Agfg1 and drongo. Two independently-produced reviews of the identical GOA
row now agree, which matters because a divergence between them would itself have been a
defect.

2. The zinc is really there, and it is held by the four predicted cysteines

zinc_site.py computes the Zn coordination shell from the deposited coordinates of both
AGFG1 ArfGAP-domain entries. Author numbering is not UniProt numbering (2D9L is
offset by +6), so the offset is read from SIFTS and each converted position is checked
against the UniProt sequence.

Two independent structures agree, and both agree with the motif scan of section 1. This
is direct experimental evidence that human AGFG1 binds zinc — a term GOA does not
carry (see section 8).

The table above is now machine-checked against zinc_site.json. The first committed
version of zinc_site.py wrote that artifact as {} — the loop computed each result,
printed it, asserted it, and never stored it — so this table had no persisted output
behind it while still validating as a byte-exact quote. audit_review.py check J parses
the rendered table back out and asserts the PDB id, offset and coordinating cysteines
agree with the JSON; check I fails on any empty artifact in this directory. Both are
break-tested against the exact content that shipped.

Neither structure has a primary publication: PDBe reports 2OLM as "ArfGap domain of
HIV-1 Rev binding protein" and 2D9L as "Solution structure of the ArfGap domain of
human RIP", both "To be published" with a null PubMed id and a null DOI. That is why
the GO:0008270 row's original_reference_id is this computation rather than a PMID.

3. AGFG1 and AGFG2 are paralogues — verified, not assumed

paralogy.py, three independent lines with a scale:

1. PANTHER: shared family PTHR46134; distinct subfamilies PTHR46134:SF1 (AGFG1)
   vs PTHR46134:SF4 (AGFG2).
2. InterPro: identical signature sets — IPR001164, IPR037278, IPR038508,
   IPR052248 — for both.
3. Pairwise identity (Biopython global alignment, BLOSUM62):

AGFG1/AGFG2 are closer than an accepted ArfGAP paralogue pair, so the shared name is
not the only reason to treat them as one clade.

4. One PANTHER node carries three AGFG1-specific spermatid terms to BOTH paralogues

node_reach.py asks the two node questions in both directions. GO:0005737 (593,028 IBA
annotations) and GO:0031410 (38,628) are not attempted — no partial count is
reported for them — because a first-page total read as a whole is exactly the pagination
trap. The three biological-process terms are small enough to enumerate fully.

The 36 recipients of PTN002919572 were counted programmatically from the WITH/FROM
fields, not by eye: 21 AGFG1-type, 12 AGFG2-type and 3 unnamed gene products across
17 species. Mouse Agfg2 (Q80WC7) and rat Agfg2 (A0A0G2K7G2) are themselves
recipients, so no AGFG2 orthologue in any species carries experimental evidence for any
of these three terms.

Meanwhile donors.py shows the entire experimental basis of all three is mouse Agfg1
(Q8K2K6) alone:

So PTN002919572 sits at or above the AGFG1/AGFG2 duplication, and AGFG1-specific
spermiogenesis biology has been attached to it. For AGFG1 the transfer is an
orthologue transfer from a strong mouse null phenotype and is sound. For AGFG2 it is
a paralogue transfer with no AGFG2 evidence anywhere. Moving the three terms down to the
AGFG1-orthologue node (PTHR46134:SF1) would retract them from the 12 AGFG2-named
recipients in one edit and leave all 21 AGFG1-named recipients correct. The 3 unnamed
recipients are not classified either way.

5. A byte-identical WITH/FROM is self-referential on AGFG1 and paralogue-derived on AGFG2

paralog_goa.py. Two rows carry UniProtKB:P52594 in their WITH/FROM:

A self-referential IBA records a PAINT curator judging the function core for that gene;
the same bytes on the sibling are a transfer from AGFG1. Same field, different evidential
status.

6. Two of the four protein-binding partners are one Y2H screen counted three times

intact.py expands the IntAct records instead of trusting NbExp.

VAMP7's seven records are one study but three orthogonal methods, one of them
quantitative. POLE2's and NYAP2's three records each are one HuRI screen logged as three
sub-methods of the same experiment — the NbExp=3 artefact — with no orthogonal assay.

The promiscuity arm of the argument comes back negative and is reported as such:
AGFG1 has 37 IntAct partners, POLE2 37 and NYAP2 8, so neither partner is a hub and
neither is topologically inaccessible to a nuclear/cytoplasmic protein. The case against
those two rows rests on the single-screen evidence alone.

7. Reference-projection test

reference_projection.py. Entity counts are distinct gene-product ids, never annotation
totals.

The PMID:11545741 result is worth stating positively: a curator annotated four
different proteins with five different terms from that paper, so the AGFG1↔NXF3 row is a
considered call rather than pipeline output — even though the paper's full text is not
retrievable (Europe PMC: isOpenAccess N, inEPMC N, no PMC id).

8. The mouse orthologue lacks all three legacy human TAS rows

Full annotation set of Q8K2K6 (17 annotations): GO:0001675, GO:0007289, GO:0045109,
GO:0005634, GO:0005737, GO:0031410, GO:0043025, and GO:0005096 (IEA:InterPro).

Mouse Agfg1 carries zero GO:0003723 rows, no GO:0005643, and no GO:0006406. All
three human rows come from the two 1995 discovery papers via ProtInc/PINC and exist in
one species only.

9. UniProt still carries a keyword-derived MF that GOA has dropped — and one it should keep

AGFG1-uniprot.txt lists both of these under DR GO:

DR   GO; GO:0003677; F:DNA binding; IEA:UniProtKB-KW.
DR   GO; GO:0008270; F:zinc ion binding; IEA:UniProtKB-KW.

Neither appears among AGFG1's 35 GOA rows — the IEA:UniProtKB-KW route was withdrawn
from GOA. The two cases differ in kind:

• GO:0003677 DNA binding is not supported. AGFG1 has no DNA-binding module. Its
  only zinc site is the ArfGAP C4 finger, which section 2 shows is buried and fully
  occupied by structural zinc, and which PMID:18809720 describes as having "an
  architectural rather than catalytic role". The KW DNA-binding keyword appears to
  follow InterPro's remark that the ArfGAP zinc finger "displays some similarity to the
  C4-type GATA zinc finger" — a fold resemblance, not a measured activity.
• GO:0008270 zinc ion binding is supported and is missing from GOA. Two structures
  resolve the zinc on four cysteines (section 2).

10. The ARBA rule that gives AGFG1 GO:0005737 fires on a gene-specific FunFam pair

parse_arba.py evaluates all 2414 condition sets of ARBA00026971 against AGFG1's
own signature/taxon profile, distinguishing sets that fire from sets that merely
mention one of its signatures.

• fires (1): FunFam id=1.10.220.150:FF:000005 AND FunFam id=3.30.450.50:FF:000005
• mentions but does not fire (1): IPR001164 AND IPR037278 AND IPR051718 — AGFG1 has
  IPR052248, not IPR051718
• the rule grants exactly one term: GO:0005737

Both FunFams in the firing set are named for AGFG1 itself, so this is a gene-specific
rule granting a location AGFG1 independently has by experiment. No defect.

11. A fusion-construct crystal structure has given AGFG1 a phantom Longin domain

Gene3D/CATH assigns G3DSA:3.30.450.50 "Longin domain" to AGFG1 residues 154–234.
That region is annotated REGION 145..193 Disordered (MobiDB-lite) and
COMPBIAS 176..191 Polar residues in AGFG1-uniprot.txt, and PMID:18775314 says the
VAMP7-binding stretch "is contained within the region that follows the ArfGAP domain,
which is predicted to lack stable secondary structural elements".

The origin is traceable. PDBe reports that each of chains A–D of 2VX8 is a single
169-residue polypeptide mapping to two UniProt entries:

P52594 AGFG1_HUMAN  chains A-D  unp 136-175
P70280 VAMP7_MOUSE  chains A-D  unp 1-120

and the paper describes making exactly that construct: "a cDNA encoding the fragment of
Hrb mapped as binding VAMP7 longin domain (residues 136–176) was cloned onto the 5′ end
of the VAMP7 longin domain cDNA". The longin fold belongs to VAMP7; the domain call has
crossed the fusion junction into AGFG1.

Negative control: AGFG2 — same family, same InterPro set, 71.2% identical over the
ArfGAP domain — has no 3.30.450.50 assignment. The error is keyed on having a
chimeric PDB entry, not on family membership. It is not currently producing a wrong GO
term (the FunFam pair grants only cytoplasm), so it is reported as a
CATH/Gene3D–UniProt correction rather than a GO action.

12. Retraction / erratum check

retractions.py checks all 20 relied-on PMIDs by two routes — CommentsCorrections
RefType on each cited article's own record, and PublicationType — with
PMID:32125225 (known retracted) as a positive control, which fires.

0/20 carry a retraction, erratum or expression-of-concern flag. A correction can
carry a null PubMed id and be visible only through Crossref, so this is "none found by
these two routes", not "none exists".

12b. Searching for an acrosome compartment found a better process term instead

Asked in review whether an acrosome-associated cellular component should be proposed.
GO:0001669 acrosomal vesicle is declined on its own definition — "A structure in
the head of a spermatozoon that contains acid hydrolases ... derived from the lysosome" —
which denotes the mature organelle, whereas AGFG1 sits on the cytosolic surface of the
precursor vesicles and in the null the mature acrosome never forms. The human evidence
is orthology-only, so a human CC row would rest on similarity for a compartment nobody has
imaged in human spermatids.

The search for that CC surfaced GO:0120211 proacrosomal vesicle fusion, whose
definition is the mouse phenotype verbatim: "Fusion of the membrane of proacrosomal
vesicle with the membrane of another proacrosomal vesicle to form the acrosome." It is a
verified descendant of the accepted GO:0001675 over is_a/part_of, so proposing it is
additive rather than a sideways move. That is a better answer to the question than the
component would have been.

13. Checks that came back negative, recorded so they are known to have been run

• Logical-opposite cross-product (a term and its negative regulation of twin
  sharing references): AGFG1's GOA contains no such pair. Not applicable.
• Promiscuity / topological-inaccessibility argument against the two HuRI partners:
  negative (section 6).
• Partner-accession substitution (TrEMBL clones or partial ORFeome constructs posing
  as canonical partners): negative. All four partners resolve to reviewed Swiss-Prot
  entries at canonical length — VAMP7 220 aa, POLE2 527 aa, NYAP2 653 aa, NXF3 531 aa.
• IBA landing above its donor: negative. Every IBA term AGFG1 receives is the same
  term the donor holds; no downward MODIFY is warranted.
• Circularity: negative. All five IBA rows have at least one donor with its own
  experimental annotation at the exact term (section 4).
• Direct evidence that AGFG1 binds RNA: a recorded Europe PMC query
  (litsearch.py, key agfg1_rna_binding, 234 hits) returned no study reporting
  RNA binding by AGFG1/hRIP. This is a statement about that query, not about the world.
• QuickGO identifier forms: probe_ids.py confirms MOD ids (MGI:MGI:1333754,
  MGI:1333754, FB:FBgn0020304, FlyBase:FBgn0020304) are rejected with HTTP 400
  while UniProt accessions return 200 — so every MOD donor was resolved to a UniProt
  accession first, with UniProtKB:P52594 (35 hits) as the positive control.

14. Recall of the affinage record

AGFG1-deep-research-affinage.md reports gates_passed: True with 2 citations
(PMID:10613896, PMID:18819912). Neither of the two papers that decide this review is
among them: PMID:18775314 (the VAMP7 structure, ITC and depletion phenotype) and
PMID:11711676 (the mouse null with no acrosome). Nor is PMID:18809720, the family
reference that supplies the catalytic-motif anchor, nor PMID:38606629, the only human
dataset placing AGFG1 with ARF1/ARF3/ARF6, nor PMID:23433073, the paper that decides the
catalytic question. A passing gate bounds precision and says nothing about recall.

And the provider is not the only thing whose recall failed here. PMID:23433073 was
fetched by this review, judged to be a family-classification paper with "nothing resting on
it", and deleted before the first commit. It contains the residue-level analysis that
reverses the review's headline verdict, and it came back only through the concurrent AGFG2
review. The lesson is the campaign's own, learned the hard way: a paper titled for the family
holds the gene's answer, and an absence created by not reading is not evidence.