AHNAK (Q09666) — review notes

Working journal for the GO annotation review. Provenance is given inline as
[PMID:xxxx "verbatim text"] or as the query that produced a number.

What AHNAK is

5,890 aa, ~629 kDa by UniProt's own calculation (the literature says "700 kDa", which is the
apparent mass on a gel). Architecture from the UniProt feature table:

• FT DOMAIN 9..90 /note="PDZ" /evidence="ECO:0000255|PROSITE-ProRule:PRU00143" — predicted, not
  experimentally solved, but see the holdup finding below, which is functional evidence that it
  really is a PDZ.
• No FT REPEAT features at all, despite KW Repeat. The tandem 128-residue central repeat
  units that the entire literature describes are not modelled in the entry. The published counts
  differ — PMID:14699089 for the 5,643-aa sequence, against PMID:10318799 — so this review says "26 to 30" rather than picking one. The
  22 FT REGION features are all MobiDB-lite "Disordered".
• Five predicted NLS motifs, all in the C-terminal region (4971-4979, 5019-5027, 5034-5039,
  5706-5716, 5772-5779).
• 13 VARIANTs, all dbSNP; the I5236T variant that PMID:16319140 characterises functionally is
  not among them.
• CC -!- FUNCTION: May be required for neuronal cell differentiation. — one hedged sentence
  for a protein with a thirty-year literature. Raised as a UniProt suggestion.
• CC -!- SUBCELLULAR LOCATION: Nucleus. — also a single word, against four plasma-membrane
  IDAs on the same protein.
• Isoform 2 (Q09666-2) = residues 1-149 with 115-149 replaced by a 22-residue alternative, i.e.
  ~136 aa = the PDZ domain alone. This is almost certainly the "small 17 kDa AHNAK" of
  PMID:21940993. Corroborating: mouse has TrEMBL entries Q8CEX7 and A0A494B8Y7 (176 aa),
  both named "PDZ domain-containing protein" / "AHNAK nucleoprotein".

Row reconciliation (done first, per campaign procedure)

wc -l AHNAK-goa.tsv        -> 67 (66 data rows)
grep -c '^- term:' stub    -> 66

66 : 66, exact. The fetch-gene under-seeding defect seen on ADAMTSL5 and ACTR5 (collapsed
GO:0005515 partner rows) is absent here — all 26 protein-binding rows were seeded
individually. Verified programmatically by matching on
(term, evidence, reference, sorted(WITH/FROM), negated) with an assertion that every TSV row
is consumed exactly once and no review key goes unused. Final file = 66 GOA rows + 9 NEW.

WITH/FROM resolution (every token)

Zero unresolved tokens. Note there is no Swiss-Prot mouse Ahnak — every Compara and CAFA row
on this gene descends from an unreviewed TrEMBL entry. Its GO annotations are curated
experimental ones; its protein name is automatic. Those are two different things (the
evidence-provenance vs name-provenance distinction).

Finding 1 — the membrane-repair annotation is a complex projection over a negative result

GO:1905686 positive regulation of plasma membrane repair, NAS, PMID:22940583, ComplexPortal.

The cited paper is a crystallography paper. Its only AHNAK content is a 3.5 Å structure of a
C-terminal peptide on the S100A10/annexin A2 heterotetramer; membrane repair appears only in the
framing sentence
PMID:22940583. NAS is the correct code for that.

Projection test (QuickGO by reference, not by gene):

reference=PMID:22940583 -> 16 annotations, 10 distinct entities, complete=True
  GO:1905686 NAS -> 10 entities   (CPX-850, CPX-853, CPX-898, CPX-905,
                                   ANXA2+S100A10+AHNAK human, ANXA2+S100A10 mouse, Ahnak mouse)
  GO:0045921 NAS ->  6 entities

All assignedBy: ComplexPortal, identical evidence code. This is the ACTR8 shape: the phenotype
term spreads across the whole set. The ACTRT3 discriminator (does the functional term stay on the
gene that was actually perturbed?) cannot even apply, because nothing was perturbed — it is a
structure paper.

And the only experiment that ever tested it is negative:
PMID:20833135

Verdict: MARK_AS_OVER_ANNOTATED, not REMOVE. The negative is one assay, in mouse skeletal
muscle, hedged by its own authors, and AHNAK's partners (dysferlin, ANXA2, S100A10) genuinely are
repair proteins. But a positive-regulation assertion with no supporting experiment anywhere and a
loss-of-function negative in the obvious assay does not deserve to stand unqualified.

Finding 2 — a plasma-membrane IDA on a paper that predates the gene

GO:0005886, IDA, PMID:3126079, ComplexPortal. That paper is Osborn et al. 1988 on p11 and p36
PMID:3126079 — i.e. S100A10 and annexin A2, with no AHNAK experiment, and it could not have one:
the AHNAK gene was first reported in 1992 PMID:1608957.

reference=PMID:3126079 -> 14 annotations, 10 entities
  GO:0005886 IDA -> 8 entities incl. UniProtKB:Q09666 AHNAK, assignedBy ComplexPortal

The term is right — plasma membrane is held on this gene by an HPA IDA and two literature IDAs
— so this is a reference defect, not a term defect. Action ACCEPT with the provenance recorded
and reference_review.correctness: MISCITED; raised for ComplexPortal as "can projections be
restricted to references postdating the subunit's identification?".

Finding 3 — one holdup screen supplies 8 of the 26 protein-binding rows

IntAct expansion of Q09666: 649 interaction records, 608 distinct partners. Of those,
415 records come from PMID:36115835 alone, every one with detectionMethod: holdup assay,
AHNAK always the prey, partners always the bait. The paper describes the design
PMID:36115835
against a library of 448 C-terminal 10-mer PDZ-binding motifs, measured by
PMID:36115835.

So those 8 GO rows report an affinity between AHNAK's isolated ~80-residue PDZ domain and a
10-residue peptide. This is the ACRV1 NbExp=3 pattern at a larger scale: one experiment
becoming eight annotations.

Judged per partner, not wholesale:

The flip side is a genuine positive: the screen is functional evidence that AHNAK's
ECO:0000255-predicted PDZ domain really folds and binds PDZ ligands.

Negative results from the sibling checks, reported because they were run:
- Partner-accession check (the ACRV1 TSC1/ORFeome trap): all 15 distinct GO:0005515
partner accessions resolve to reviewed canonical Swiss-Prot entries at full length — the HCV
row correctly uses the mature-chain id P27958-PRO_0000037576 rather than a truncated clone. No
ORFeome-fragment substitutions. The fragment on these rows is on AHNAK's side, not the partner's.
(This claim is about partner accessions only. Several WITH/FROM donor accessions are
TrEMBL, not Swiss-Prot — mouse Ahnak E9Q616 and all three mouse Ahnak2 entries — as the
table above records. Conflating the two sets would have made a false 19-of-19 claim.)
- Topological-impossibility check: mostly negative. HSPA5/BiP is ER-lumenal against a
cytosol-facing AHNAK, which is worth noting, but the other partners are all cytosol- or
nucleus-accessible, so the ACRV1-style "every partner is on the wrong side of a membrane"
argument does not apply here.

Finding 4 — GO:0043484 is one experiment reaching 74 proteins by three routes

Donor-evidence query for GO:0043484 across the family:

So the only experimental annotation to this term anywhere in the family is the single mouse
IDA. PAINT node PTN001156025 then gives it to 74 gene products, and human AHNAK receives it
three times — IBA, Ensembl Compara IEA, and CAFA ISS — from that one experiment.

Two mitigations that make the PAINT call more defensible than it first looks:
1. The paper tests both family members: PMID:21940993 So the
node has two members behind it, not one. (Corollary: Periaxin's experimental result was never
captured as an experimental annotation — human/mouse/rat PRX all hold the term by IBA/ISO only.
Raised for PAINT.)
2. The node's placement is clean — PTN001156025 gives exactly GO:0005634, GO:0005737,
GO:0043484 to exactly the same 74 entities, and all three human members get exactly the same
three terms. No AADACL-style node-misplacement defect here; reported as a negative.

What is wrong is the level: the activity belongs to the short nuclear isoform (Q09666-2), not
to the 700 kDa protein the canonical accession represents. All three rows → KEEP_AS_NON_CORE.

Finding 5 — the CAFA ISS block cites one paper for five different terms

Five ISS rows (GO:0005634, GO:0005829, GO:0043034, GO:0043484, GO:0097493), all citing
PMID:21940993, all donor UniProtKB:E9Q616. But the donor's own evidence for two of them is in
a different paper:

E9Q616 -> GO:0043034 : IDA PMID:20833135     (costamere / vinculin colocalisation)
E9Q616 -> GO:0097493 : IMP PMID:20833135     (AFM fibre stiffness)
E9Q616 -> GO:0005829 : IDA PMID:21940993     (agrees with the row)
E9Q616 -> GO:0043484 : IDA PMID:21940993     (agrees with the row)

PMID:21940993 is about alternative splicing at the AHNAK locus and contains neither the costamere
nor the stiffness experiment. Block-assigned ISS rows carrying one reference for the whole block is
the mechanism. Raised for UniProt/CAFA.

Finding 6 — an MF term inferred from whole-fibre mechanics

GO:0097493 structural molecule activity conferring elasticity — definition: "The action of a
molecule that contributes to the structural integrity of a complex or assembly within or outside a
cell, providing elasticity and recoiling."

Its entire evidence is one mouse IMP: PMID:20833135

Care with direction — I initially wanted to call this an inversion and it is not: losing AHNAK
making fibres stiffer is consistent with AHNAK providing compliance. The objection is about
level, not sign: a whole-fibre knockout phenotype is being read as a molecular property of the
protein. No force spectroscopy, extensibility measurement or reconstitution on AHNAK or its repeats
has ever been published. Both rows → MARK_AS_OVER_ANNOTATED; the experiment that would settle it
is in suggested_experiments.

Finding 7 — the T-tubule contradiction that dissolved on checking

Initial read: GO:0030315 T-tubule NAS from PMID:17185750 ["In normal skeletal muscle, dysferlin
and AHNAK colocalize at the sarcolemmal membrane and T-tubules."] is flatly contradicted by
PMID:20833135, and the
later paper used AHNAK1- and AHNAK2-specific antibodies while the earlier did not — a textbook
paralogue cross-reactivity artefact.

Then I checked a third paper and the story changed. PMID:12153988 The disagreement is
confined to skeletal muscle; in human cardiac muscle the localisation is independently
supported. So it is a tissue difference, not a contradiction, and the honest action is ACCEPT
with both observations recorded — not the MARK_AS_OVER_ANNOTATED I was heading for.

Logged as a reminder: the campaign's over-annotation bias nearly produced a confident wrong answer
here, exactly as the AADACL4 "a general term can be correct" note warns.

The real diagnosis: this is a coverage defect, not an over-annotation defect

Aspect breakdown of the 66 GOA rows:

• CC: 31 rows, 16 distinct terms. Plasma membrane ×5, nucleus ×4, exosome ×3, cytosol ×3,
  cytoplasm ×3, costamere ×2, sarcolemma ×2, plus plasma membrane protein complex, membrane raft,
  cell-cell contact zone, vesicle, T-tubule, membrane, actin cytoskeleton, focal adhesion,
  lysosomal membrane.
• MF: 31 rows, 5 distinct terms — and 26 of the 31 are bare GO:0005515 (21 assigned by
  IntAct, 4 by UniProt, 1 by AgBase). The other four terms are GO:0044548 ×1, GO:0045296 ×1,
  GO:0003723 ×1 and GO:0097493 ×2, the last of which this review marks over-annotated.
• BP: 4 rows, 2 distinct terms — GO:0043484 ×3 (all three routes out of one experiment) and
  the unsupported GO:1905686 ×1.

(Counted with awk -F'\t' 'NR>1 {print $7}' AHNAK-goa.tsv | sort | uniq -c. My first pass
hand-counted these off a printed row list and got 39/23/4 with 17/4/2 distinct — wrong in five
places. Derive the number, do not eyeball it.)

So a protein with hundreds of papers has effectively zero curated biological process and no
informative molecular function at all. The things AHNAK is actually known for are absent:

The last row was nearly missed. PMID:15177040 entered this review only as the source of the
"weak DNA-binding activity" caveat on the GO:0003723 row — cited, but supporting no annotation.
Re-reading it for that caveat surfaced the actual result: PMID:15177040,
with purified proteins and an in-cell co-IP. A cited reference that supports no annotation is a
flag, in the same way the brief flags an empty findings: list on a full-text reference.

Term choice was constrained: GO:0051106 positive regulation of DNA ligation is obsolete
(checked at QuickGO, not inferred from an empty search), and involved_in GO:0006303 would
over-reach because no cellular repair assay exists. GO:0008047 enzyme activator activity —
"A molecular function regulator that increases a catalytic activity" — is what was measured.

Hence 9 NEW rows. The two MODIFYs move bare GO:0005515 S100A10 rows to GO:0044548, which the
gene already carries — the only partner on this protein with crystal structures behind it.

GO:0005515 on DYSF/myoferlin was left as-is deliberately: it is the strongest IPI on the gene
(direct, Ca-independent, domain-mapped both ways) but GO has no term for binding a ferlin, and
GO:0019904 protein domain specific binding would be true and no more informative. Filed as an
ontology question rather than forced.

Reference hygiene

• Retraction/erratum sweep over all 58 PMIDs touched, reading CommentsCorrections/RefType
  from each cited article's own PubMed record (a Publisher Correction is invisible to a
  publication-type search): 0 retractions, 5 errata — 36115835→36477203,
  39251607→39468017, 26466345→26629899, and 1608957 and 15489334 with errata carrying
  no PMID. All are Author Corrections; none has an abstract, so scope could not be established.
  Recorded in the relevant reference_review blocks. Nothing load-bearing rests on them.
• affinage: gates clear, 45 numeric-PMID citations, no PMID:bio_* preprint ids. Spot checks
  found it accurate except that it describes PMID:26466345 as BMP2-driven where the paper reports
  BMP4/SMAD1. It supplied PMID:20833135, this review's decisive reference, which appears
  nowhere in AHNAK's GOA reference set — a recall win worth recording. Per campaign policy no
  affinage sentence is quoted as supporting_text; every claim it contributed was re-grounded on
  the primary paper. This is why just validate reports the "no annotations reference available
  deep research files" warning, which is left standing deliberately.
• All 87 supporting_text entries verified (85 literature via the repo's own
  SupportingTextValidator, 2 file: quotes by literal substring, since CI does not check
  file: quotes). Checker self-tested by breaking it three ways: a mutated literature quote, a
  fabricated file: quote, and a file: quote spanning a UniProt CC continuation line — all
  three fired.

A number that refused to add up

First dump: raw supporting_text keys=68 against parsed quote entries=87. Resisting the urge to
find a story, the gap traced to PyYAML anchors and aliases — shared Python dicts became
&id001/*id001. Accounting: 14 aliased entries from the two shared holdup quote dicts (8 uses),
2 from the shared splicing quote dicts, 2 from the shared splicing list, 1 from the shared
elasticity list = exactly 19 = 87 − 68. copy.deepcopy does not fix this (it preserves
internal sharing by design); a dumper with ignore_aliases does. The builder now asserts no
anchor survives and that raw == parsed.

What I did not do

• No bioinformatics analysis. The claims in this review rest on published measurements and on
  database queries (QuickGO, IntAct, UniProt REST) whose commands are recorded above; nothing here
  needed a new alignment or structure calculation.
• Did not read the full text of PMID:21423176, PMID:25468996 or PMID:36115835 beyond the
  parts quoted; AHNAK appears in the first two only in supplementary tables. Those rows are
  KEEP_AS_NON_CORE accordingly.