AGFG1 (P52594, HRB / RIP / RAB) — review notes

Working journal for the GO annotation review. Computed results live in
AGFG1-bioinformatics/RESULTS.md; this file records the reasoning, the process
history, and the leads that did and did not pay off.

What the gene is

562-aa protein with an N-terminal ArfGAP domain (UniProt DOMAIN 11..135 Arf-GAP,
ZN_FING 29..52 C4-type) followed by ~430 residues of low-complexity sequence carrying
ten phenylalanine–glycine (FG) repeats, three FxxFxxF AP-2-appendage motifs and NPF
motifs. Two names, two literatures: the 1995 HIV work (hRIP/Rab/RIP — a Rev cofactor)
and the 2001–2008 trafficking work (Hrb — an endocytic clathrin adaptor and an acrosome
biogenesis factor).

PE 1: Evidence at protein level, so the protein is detected; what is thin is
biochemical characterisation of the ArfGAP activity, not detection.

Row-count reconciliation (do this first)

AGFG1-goa.tsv          36 lines = 35 annotations + header, 35 distinct
AGFG1-ai-review.yaml   34 '- term:' entries as seeded

The stub collapsed the two GO:0005515 / IPI / PMID:25416956 rows, which differ only in
the WITH/FROM partner (UniProtKB:P56282 POLE2 and UniProtKB:Q9P242 NYAP2). Restored
to one entry per partner, so the review has 35 entries against 35 distinct GOA rows.
This is the documented GOAValidator.seed_missing_annotations behaviour (the seed key
omits WITH/FROM), not corruption — but it hides exactly the granularity that per-partner
judgement needs, and here the two partners get the same verdict only because they came
from the same screen.

The paralogue question, answered rather than assumed

The task brief said "the shared name implies they are paralogues — verify that". Done
three ways in paralogy.py: shared PANTHER family PTHR46134 with distinct subfamilies
SF1/SF4, identical InterPro signature sets, and 71.2% identity over the ArfGAP domain
(48.9% full length) against 48.2% for the accepted ARFGAP1/ARFGAP3 paralogue pair and
19.6% for an unrelated control. They are genuine, relatively recent paralogues.

The finding: one node, two paralogues, one gene's evidence

PTN002919572 carries GO:0001675, GO:0007289 and GO:0045109 to 36 gene
products, whose human reach is exactly {AGFG1, AGFG2}. All three rest on mouse Agfg1
IMPs and nothing else. Mouse and rat Agfg2 are themselves IBA recipients, so no AGFG2
orthologue in any species carries experimental evidence for any of the three.

For AGFG1 this is an orthologue transfer from a decisive mouse null
PMID:11711676, so the rows stand. For AGFG2 it is a
paralogue transfer. And AGFG2 has its own, unrelated characterised function — knockdown
in HUVECs halves stimulated von Willebrand factor secretion from Weibel–Palade bodies
PMID:34369554,
a paper that does not mention AGFG1 at all. So the divergence is real and the node's
reach is too broad.

Recommendation filed in suggested_questions, stated once with both genes named:
move the three terms from PTN002919572 to the AGFG1-orthologue node
(PTHR46134:SF1). Relayed to the concurrent AGFG2 review as a claim to verify, not
as a fact.

The reciprocal half

Two of AGFG1's own rows (GO:0005737, GO:0031410) carry UniProtKB:P52594 in
WITH/FROM. On AGFG1 that is self-referential — a PAINT curator judging the function
core, which is valid. The byte-identical field on AGFG2 makes those rows
paralogue-derived. Same bytes, different evidential status; this is the AFF1/AFF4
pattern from the campaign brief, seen from the donor side.

The "GAP domain in the name" lead: CONFIRMED after all — my first answer was wrong

Retraction, recorded rather than quietly overwritten. The section below originally
concluded that the pseudoenzyme hypothesis was NOT confirmed, because the arginine finger
is intact. That was an artefact of testing one of the three residues the field says
Arf GAP catalysis requires. AGFG1 retains one of three. GO:0005096 moved from
KEEP_AS_NON_CORE to MARK_AS_OVER_ANNOTATED. The reasoning below is preserved because
it is still correct about what it tested and it is instructive about how the error
happened; the correction is in the subsection that follows it.

How the error happened, precisely: I anchored on PMID:18809720, which names the arginine
and the four cysteines, and treated "the apparatus" as exhausted by what that paper names.
The decisive paper — PMID:23433073 — I had fetched and then deliberately deleted,
reasoning that "nothing rests on it" because it looked like a family-classification paper.
It contains the residue-level analysis that decides the question. The campaign brief's own
warning applies exactly: a paper titled for the family holds the gene's answer, and an
absence I created by not reading is not evidence.

It was surfaced by the concurrent AGFG2 review, which is the strongest argument for
reviewing paralogues in parallel that this gene produced. I verified it independently
rather than inheriting it — see the subsection below — and the verification reproduced the
sibling's AGFG2 number (Thr89) exactly while adding AGFG1, mouse Agfg1 and drongo.

The original reasoning, preserved: the arginine finger is intact and the ADAP control matters

The brief flagged PSEUDOENZYME_OVERANNOTATION as the shape to look for. ~~It is absent
here~~ — this sentence is the retracted claim; it is present here, see the correction
below — and the check is still worth recording because the anchor is published rather than
remembered: PMID:18809720 gives the ArfGAP catalytic spacing as
PMID:18809720
and names ArfGAP1's residues as Cys22/25/42/45 + Arg50. arfgap_motif.py reproduces
those numbers as a positive control and then finds AGFG1 intact at Cys29/32/49/52 +
Arg57. Both AGFG1 structures resolve the zinc on exactly those four cysteines
(2.26–2.40 Å), computed from coordinates with the SIFTS numbering offset handled.

But the same paper supplies the disconfirming control for the opposite over-reading:
PMID:18809720 — and ADAP1's CX2CX16CX2CX4R is complete. An intact apparatus is necessary
and demonstrably not sufficient. So the motif result cannot be used to certify
GO:0005096 either.

Round 2 of review caught that I was using half of that sentence. Its first clause is a
blanket positive covering the AGFG subfamily — the paper puts the 31 human ArfGAPs into
ten subfamilies, AGFG is one, and only the ADAPs are excepted — so quoting the ADAP half as
a control while passing over the rest is the "verbatim but selectively bounded" defect, and
no gate can see it because the quote is verbatim. Both clauses are now argued, with four
checkable grounds for discounting the blanket clause's AGFG instance: it carries no
citation while the sentences either side of it do; it is stated at subfamily level and names
no species, protein or assay; the paper's own AGFG section supplies no measurement; and the
same paragraph warns that PMID:18809720, which is precisely what a subfamily-level summary cannot
settle. Net effect on the verdict: none — but it does mean the clause raises the prior of
catalytic competence, which is part of why the row is kept rather than marked
over-annotated.

What tips GO:0005096 to "keep, non-core" rather than "over-annotated":

That was the reasoning for KEEP_AS_NON_CORE, and it is superseded by the measurement below.

The correction: two of three required residues are gone, subfamily-wide

PMID:23433073 names three catalytically required positions in ASAP3 and states
PMID:23433073. R469 is the arginine finger; D484 contacts the Arf6 catalytic glutamine Q67 and
stabilises switch 2; W451 sits in the Arf–ArfGAP interface.

Measured in catalytic_residues.py with three controls that must all pass first — ASAP3's
own numbering still holding, every GAP-competent panel member recovering all three, and
the arginine agreeing with the independent motif scan (8/9 entries):

protein W451 R469 D484 n/3
ASAP3, ARFGAP1, ARFGAP3, ASAP1, SMAP1 W R D 3/3 each
AGFG1 human Y39 R57 T71 1/3
AGFG2 human Y57 R75 T89 1/3
Agfg1 mouse Y39 R57 T71 1/3
drongo Y40 R58 A72 1/3

The source paper predicts exactly this from 40 AGFG sequences
PMID:23433073 and
PMID:23433073, and concludes
PMID:23433073.

The reversal resolves the tension in my own evidence rather than creating one. The ARF
proximity result had looked like support for catalysis; the same paper says it is not
PMID:23433073. Binding retained + catalysis lost = an Arf effector. And the fly
genetics is reinterpreted rather than discarded: drongo scores 1/3 too, and an effector that
antagonises an Arf-GEF without hydrolysing GTP fits that genetics as well as a GAP does.

MARK_AS_OVER_ANNOTATED rather than REMOVE: the domain is genuine, the zinc is really
bound, Arf binding is real, and no assay has been run on the human protein in either
direction.

Note GO has merged the substrate-specific GAP terms (GO:0008060 and eight others are
secondaryIds of GO:0005096), so GO:0005096 is already maximal and no substrate-
specific child should be proposed.

The three legacy TAS rows are the real defects

All three come from the two 1995 discovery papers via ProtInc/PINC, each annotating
exactly one entity in all of GOA, and none exists on the mouse orthologue.

GO:0003723 RNA binding → REMOVE. The direction of the interaction is inverted. The
cited paper has Rev binding the RNA and hRIP/Rab binding Rev's protein activation
domain PMID:7634337, and the
Rev–AGFG1 contact is not even direct PMID:18809720. AGFG1 has no
RNA-binding module: the only nucleic-acid-suggestive feature is the ArfGAP C4 zinc
finger, which is buried and structural, and the FG repeats are protein-interaction
motifs that bind CRM1 and EH domains. UniProt records only a hedge — FUNCTION: ... May play a role in RNA trafficking or localization. — and assigns no RNA-binding keyword.
Mouse Agfg1 has zero GO:0003723 rows. A recorded Europe PMC query found no study
reporting RNA binding by AGFG1 (that is a statement about the query, not the world).
Same shape as the ADIPOQ sialic-acid case: a modification/partner study read as a
binding activity.

GO:0005643 nuclear pore → REMOVE. The term's definition is complex membership — "A
protein complex providing a discrete opening in the nuclear envelope" — and the cited
paper claims only sequence resemblance PMID:7637788. A primary
paper says so outright PMID:10613896. UniProt, curating the same two papers, records Nucleus and
Cytoplasmic vesicle and neither nuclear envelope nor nuclear pore. AGFG1 is not a
nucleoporin by any signature.

GO:0006406 mRNA export from nucleus → MODIFY. The row is not false so much as
unable to say what it needs to: the evidence is entirely Rev/RRE-dependent viral RNA,
and the same lab later showed host mRNA is untouched PMID:14701878. The obvious
viral term cannot be used: GO:0046784's definition and all its synonyms specify
intronless viral mRNA, while Rev exports unspliced and partly spliced (intron-
retaining) RNA. So the row goes to GO:0075733 intracellular transport of virus, whose
definition ("directed movement of a virus, or part of a virus, within the host cell") is
satisfied by what was actually measured, and the missing term is filed under
proposed_new_terms. This is a sideways move between the host-process and
viral-process branches, not a generalisation — GO:0075733 is verified not to be a
descendant of GO:0006406.

Protein-binding rows, decided per partner

Cytosol: sixteen rows, one claim

Sixteen GO:0005829 TAS rows differ only in the Reactome reaction id. One
(R-HSA-8863718, "AGFG1 binds VAMP7") is about AGFG1; the other fifteen are generic
downstream clathrin-mediated-endocytosis steps — BAR proteins binding the pit, SNX9
recruiting actin machinery, synaptojanin hydrolysing PI(4,5)P2, dynamin scission,
HSPA8-mediated uncoating — in which AGFG1 has no described role. They are pathway
membership carried through every reaction, not sixteen independent statements. Each gets
its own entry so the row count reconciles, all with the same verdict and a shared
analysis: KEEP_AS_NON_CORE. AGFG1 does have a cytosolic pool from which it is recruited
to coated pits, so the term is not wrong; it is just the least informative of its
locations.

What GOA is missing (the coverage side)

GO:0035615 clathrin-cargo adaptor activity, GO:0072583 clathrin-dependent endocytosis, GO:0005905 clathrin-coated pit, GO:0030136 clathrin-coated vesicle and
GO:0008270 zinc ion binding are all supportable and none is in GOA. The 2008 Cell paper
that establishes the mechanism has produced exactly two annotations in all of GOA,
both bare GO:0005515 (on AGFG1 and on VAMP7), from a study with a structure, an ITC
constant, reciprocal mutagenesis and a depletion phenotype. That is an under-curation
gap of the AFF4 kind rather than an over-annotation problem.

One restraint worth recording: I did not propose GO:0030276 clathrin binding. The
paper calls Hrb a clathrin adaptor "in that it can be recruited into forming CCVs by its
interactions with AP2 appendages and clathrin (Figure S1 and Schmid and McMahon, 2007)" —
Figure S1 is the sequence figure and the citation is to a review, so the clathrin and
AP-2 contacts are motif-based inference in that paper, not measurements. The functional
term GO:0035615 is supported by cargo binding plus colocalisation plus the depletion
phenotype; the direct clathrin contact is not.

Also not proposed: anything from PMID:39089666 (AGFG1 raising cholesterol biosynthesis
via CAV1 in pancreatic cancer cells) or PMID:15749819 (hRIP required for HIV-1
replication). Both are real human results absent from GOA, but one cancer-cell study is
not a basis for new process terms here. Noted as knowledge gaps instead.

An unresolved contradiction in the literature, left unresolved

Two 2008 papers disagree about how general AGFG1's endocytic role is.
PMID:18819912 argues for a general clathrin-dependent-endocytosis
requirement; the other reports the opposite for a different cargo — Pryor et al. found
that depleting Hrb "had no effect on the internalization and intracellular accumulation
of EGF nor on its degradation", which they read as evidence for a SNARE-specific
adaptor. Both are recorded; the review annotates GO:0072583 clathrin-dependent endocytosis (which both support) and does not adjudicate the generality. The affinage
narrative states the general version as settled, which it is not.

The Longin-domain artefact (a database defect, not a GO defect)

Gene3D assigns G3DSA:3.30.450.50 "Longin domain" to AGFG1 residues 154–234 — a region
UniProt annotates as disordered and polar-biased. The cause is traceable: PDB 2VX8's
chains are a fusion of AGFG1 136–175 onto the mouse VAMP7 longin domain, exactly as
the authors describe building it, and the fold call has crossed the junction into AGFG1.
AGFG2, at 71.2% identity over the ArfGAP domain and with the identical InterPro set, has
no such assignment — so the error is keyed on having a chimeric PDB entry, not on family
membership. It currently produces no wrong GO term (the derived FunFam pair is one
condition of ARBA00026971, which grants only cytoplasm), so it is filed as a
CATH/Gene3D–UniProt correction request rather than acted on in GO.

UniProt correction requests

Process history