AGGF1 (Q8N302) — review notes

The question this gene poses

The symbol Angiogenic factor with G-patch and FHA domains 1 is a function claim
wrapped around two domain claims, and the three have to be adjudicated
separately. They come out differently:

claim verdict
angiogenic factor supported by a large, mostly organismal literature — but that is a BP and a receptor-ligand MF, not what the rest of the protein does; and the human-genetic evidence that named it is now retracted in substance
G-patch domain intact and canonical (8/8 conserved columns). The derived GO term GO:0003676 nucleic acid binding is nevertheless a domain-name mapping that no G-patch protein anywhere holds experimentally
FHA domain every phosphothreonine-anchoring residue retained (G437-R438, S454, N478) — and not one phosphopeptide has ever been measured. Untested, not refuted

Domain architecture — an RNA-processing protein wearing a vascular name

From AGGF1-uniprot.txt (714 aa, PE 1: Evidence at protein level):

OCRE + FHA + G-patch is the architecture of a nuclear RNA-processing factor. The
angiogenic activity maps to a ten-residue linker motif between the two named
domains. That is the shape of the whole problem: the name attributes the
function to the domains, and the mapping says otherwise.

The human genetics that named the gene does not hold up

This matters because it is the origin of every "angiogenic factor" framing, and
because UniProt still presents the variant as functional.

Provenance for those two figures, which are the only quantities in this
review with no script and no cached quote behind them: the gnomAD numbers were
read from the gnomAD v4 GraphQL API for variant 5-77035624-G-A (GRCh38) and
cross-checked against Ensembl REST /variation/human/rs34203073?pops=1; the
ClinVar classification from NCBI eutils esummary for VCV000402345. Neither
is reproduced by a committed script — flagged here rather than left implicit.

Allele frequency is not carrier frequency, and an earlier draft of the
description conflated them. At AF = 0.0142 the carrier frequency is
2·AF·(1−AF) = 2.8%, which is what the primary source actually measured:
Barker et al. found the change in 3.3% of 275 controls (9/275 = 3.27%). The
three numbers agree; the description now names the right one.

Correction to report to UniProt. FT VARIANT 133 /note="E -> K (displays a stronger angiogenic activity; dbSNP:rs34203073)" presents a benign common
polymorphism as a functional disease variant. The in vitro CAM-assay difference
Tian et al. measured may well be real; the disease attribution is not.

What this does and does not touch. It does not refute the biochemistry —
purified AGGF1 promotes angiogenesis in the CAM assay, binds endothelial cells,
and drives proliferation regardless of what E133K means. It does mean that
"AGGF1 is the Klippel-Trenaunay gene" should not be stated flatly, and that no
GO annotation should be propagated from the KTS association.

The G-patch: intact domain, unmeasured activity, misdirected GO term

Measured (see AGGF1-bioinformatics/RESULTS.md): AGGF1 matches 8/8 of the
columns conserved at ≥80% across 34 annotated G-patch domains in 33 reviewed
human proteins, including the invariant Gly at 631 and 639 (100% of the panel).

The InterPro2GO consequence is GO:0003676 nucleic acid binding from
InterPro:IPR000467. Querying the whole reference class settles what that term
is worth: 11 of 11 reviewed human G-patch proteins carry GO:0003676 by IEA
and by nothing else; zero carry it with any experimental code; six carry
GO:0003723 RNA binding (or a child) experimentally. The term is the
mapping's uniform output, not anybody's finding — and where it has been tested,
the answer is the child term.

This is reinforced by what G-patch domains are now understood to be. The
crystal structure of DHX15 bound to the NKRF G-patch shows the motif tethering
the helicase's catalytic core to its flexibly attached C-terminal domains:
PMID:32179686
and PMID:32179686
It is a helicase-activation peptide, not an autonomous nucleic-acid-binding
module.

For AGGF1 specifically, RNA binding has been measured directly — RNA
immunoprecipitation against NEAT1 in human cells, PMID:35608889 So the IEA's conclusion survives on
independent grounds, at the more informative term. Hence MODIFY rather than
REMOVE.

Helicase cofactor status, stated carefully. GOA carries an AGGF1–DHX15
interaction, and DHX15 is the canonical G-patch client. But the evidence is a
single HuRI Y2H (PMID:32296183) logged as three sub-methods; a peer-reviewed 2026
survey of the family still lists AGGF1's interacting helicase as Unknown
(PMID:42052570, Table 3). A bioRxiv preprint posted 2026-06-30
(doi:10.64898/2026.06.30.735655, Memet et al.) reports that AGGF1's isolated
G-patch stimulates DHX15 ATPase and tightens its RNA binding — unrefereed, no
PMID, and therefore not used as evidence anywhere in this review. Recorded as a
watch item and as a suggested experiment.

The FHA: every anchor residue retained, no phosphopeptide ever tested

Anchors from CHEK2 and Rad53 FHA1, verified against the live sequences and
transferred by domain-local alignment, all land inside AGGF1's annotated FHA:
G437, R438, S454, N478. The two anchor sets agree exactly wherever both
resolve. (A first attempt aligned full-length sequences globally and mapped every
CHEK2 anchor into AGGF1 residues 112–169 — a clean-looking "the FHA is
degenerate" result that was pure alignment noise. The script now refuses to align
anything but the domain windows, and the docstring records why.)

The mutational data point at the same surface. PMID:33069768 tested six COSMIC
somatic FHA variants: G437E, Q467H, Y469N and N483T impair AGGF1's
tumour-suppressor activity, while R447Q and V497I do not. G437 is the
conserved Gly of the pThr-anchoring G-R pair; R447 is not the conserved Arg
(that is R438) and V497 lies outside the 434–487 core.

And yet: no phosphopeptide, phosphothreonine or phospho-dependence experiment
on AGGF1's FHA exists. Both published FHA functions are plain co-IP with
deletion constructs —

— neither with a phosphatase control, a phospho-site mutant, or peptide
competition. There is also no experimental structure of any part of human AGGF1
(Q8N302 carries zero PDB cross-references). So no phosphoprotein-binding term
is proposed; the hypothesis is filed as an experiment.

Worth flagging as odd: 14-3-3 is itself the canonical pSer/pThr reader, so "FHA
domain binds 14-3-3" is a mechanistically unusual claim in exactly the direction
that a phosphatase control would test.

The nucleus is missing from GOA entirely

AGGF1's GOA localisation record is cytoplasm (×2), extracellular region (×3) and
perinuclear region of cytoplasm (×1). There is no nucleus annotation at all,
against three independent lines:

  1. PMID:35608889
  2. PMID:33471274 — headed by Tian XL, the first author of the 2004 discovery paper, so a different group but the same lineage; maps an NLS to residues 260–288
  3. AGGF1's own interaction record: three of the four HuRI partners — DHX15, MCRS1 and FBXO28 — carry
    GO:0005634 nucleus with experimental evidence (EXP or IDA), while the
    fourth, MAB21L3, has no cellular-component annotation in GOA at all and
    no UniProt SUBCELLULAR LOCATION comment; and AGGF1 was included as a
    spliceosome-associated protein in the Hegele 2012 Y2H matrix (PMID:22365833)

The interaction rows in AGGF1's own GO record already point at the nucleus while
its localisation rows do not.

Is "secreted" solid?

Yes, on more than one lab's data, and worth stating because there is no signal
peptide. Tian 2004 used competitive ELISA on matrigel-culture media from
untransfected HUVECs plus ³⁵S metabolic labelling, and was explicit about the
route: PMID:14961121. Independently, AGGF1 protein is measurable in human vitreous fluid
(PMID:39905000), and purified extracellular AGGF1 engages a cell-surface
receptor (PMID:34551592). The extracellular rows stand; the mechanism of export
is still unknown, which is a legitimate ontology-free gap rather than a
curation error.

Per-partner judgements on the binding rows

Nine GO:0005515 rows and three GO:0042802 rows, from seven partners. They do
not all deserve the same verdict (AGGF1-bioinformatics/RESULTS.md §4):

A live contradiction in the literature worth flagging to curators

Two co-IP-based papers disagree on the direction of AGGF1's effect on the
TWEAK–Fn14 axis:

i.e. AGGF1 blocks TWEAK–Fn14 in atrophic muscle and promotes it in retinal
endothelium. Both may be context-specific; neither is currently annotated. No GO
term is proposed for either until it is resolved.

One laboratory has done almost all of this work

Raised by the PR reviewer and then measured rather than argued
(AGGF1-bioinformatics/lab_independence.py, which reads the senior author out of
each cached record):

Over the 27 cached AGGF1 primary studies reachable from this review's own
inputs (GOA references + affinage citations + review references, minus the
high-throughput screens and the non-studies):

senior author papers
Wang Q / Wang QK 18 of 27
trainee/collaborator lineage — senior author is themselves an author on a Wang-group paper 5 — Tian XL (23628701, 33471274), Xu Y (26850475), Zhou B (28958996), Lu Q (32061268)
senior author on no Wang-group paper 4 — Qi J (17884784), Zhang JH (29885663), Liao S (33168501), Chen L (39905000)

Three of the four are used. PMID:29885663 (J Neuroinflammation 2018) corroborates
the PI3K/AKT axis, in rats; PMID:39905000 (Nature Communications 2025)
independently re-confirms the TNFSF12 interaction and measures AGGF1 in patient
vitreous; and PMID:33168501 (Liao S) shows AGGF1 co-localising with γH2AX in
human HCT116 cells, which is independent human-cell corroboration that AGGF1 is
nuclear — cited on the GO:0005634 row for localisation only, not for its
DNA-repair interpretation. Only PMID:17884784 (Qi J, a 2007 yeast two-hybrid
Id1 report) is not used.

An earlier version of this section applied a significance filter that named
only the first two and left PMID:33168501 out. That filter was a judgement
layered on top of a measurement — the count is derived end to end, and then a hand-applied
significance filter quietly dropped a paper that bears directly on this review's
largest proposal. The reviewer caught it; the lesson is the same one as the
hand-written paper list two rounds earlier, one level up.

The first version of this table under-counted the independent set, and it was
wrong — because lab_independence.py used a hand-written paper list that had
silently omitted PMID:29885663. A list you curate is a list you can under-curate,
and the conclusion then describes the list rather than the literature. The script
now derives its paper set and the lineage test computes "is this senior author an
author on a dominant-group paper" instead of only checking the discovery paper's
first author.

One caveat the script reports rather than hides: Xu C appears on PMID:39905000
and on 14 of the dominant group's 18 papers. A shared surname-plus-initial is
not proof of the same person, the senior author differs, and the affiliations are
a different institution — so the paper is still counted as independent. The
overlap is printed because the independence claim is load-bearing here.

This matters for how the review is worded, not for any action. An earlier draft
credited the nucleus evidence to more distinct laboratories than the author lists
support, and hedged the paraspeckle paper by anchoring the nucleus proposal on
PMID:33069768 "instead" — but that paper has the same senior author, so the
hedge was empty. Both statements are withdrawn. (The retracted wording is
paraphrased rather than quoted here on purpose: audit_claims.py greps for it,
and a guard that has to tell a retraction apart from a report of one is a guard
that will eventually let the claim back in.)

The useful version: AGGF1's mechanistic literature is overwhelmingly one group's
programme, with three independent contributions the review actually uses —
PMID:39905000 (TNFSF12–FN14, and AGGF1 in patient vitreous), PMID:29885663
(PI3K/AKT, in rats) and PMID:33168501 (nuclear co-localisation with γH2AX in human
HCT116). None of them touches the paraspeckle or splicing results, which remain
entirely single-group; independent replication of those is the most valuable thing
anyone could do for this gene's annotation.

A review that collapses NbExp into independent experiments and catches a
bait-labelling artefact has no business asserting laboratory independence it never
checked. The check is now committed so the claim stays a measurement.

A fetcher defect to report, not a curation one

publications/PMID_33168501.md has title: '[ole of AGGF1 in DNA damage repair…'
— the R of [Role has been eaten, presumably by a bracket-stripping step in
the publication fetcher. Translated PubMed titles are conventionally wrapped in
square brackets, so this would affect every record of that form, not just this
one.

The review's references[].title reproduces the cached string verbatim, which is
what keeps title validation passing. Hand-correcting it in the YAML alone would
mask the bug, so it is recorded here instead, for whoever owns
src/ai_gene_review/etl/publication.py.

Process notes