AFF3 — computed evidence for the GO review

This file is hand-written, not generated. It records and interprets the output of the
committed scripts in this folder; the machine-readable outputs (withfrom_resolution.json,
reference_projection.json, corrections.json) are the artifacts a re-run reproduces. Writing
it by hand deliberately avoids the regeneration trap where a hand-edit to a generated report is
silently reverted by the next run — the JSON is the claim, this prose interprets it.

Reproduce with:

uv run python genes/human/AFF3/AFF3-bioinformatics/resolve_withfrom.py
uv run python genes/human/AFF3/AFF3-bioinformatics/reference_projection.py
uv run python genes/human/AFF3/AFF3-bioinformatics/term_relations.py
uv run python genes/human/AFF3/AFF3-bioinformatics/corrections_check.py
uv run python genes/human/AFF3/AFF3-bioinformatics/intact_partners.py
uv run python genes/human/AFF3/AFF3-bioinformatics/audit_claims.py --self-test
uv run python genes/human/AFF3/AFF3-bioinformatics/verify_file_quotes.py

fix_intact_counts.py, fix_sign_claim.py, fix_pmid_count.py, apply_review_round1.py and
apply_review_round2.py are one-shot repairs, already applied. They are committed so that every correction this review made
to itself is reproducible and auditable rather than an untraceable hand-edit, and because each
one records in its docstring what was wrong and how the wrong version arose. Each asserts its
anchors are present before replacing, re-greps afterwards, and asserts detected == changed.

1. WITH/FROM resolution and donor evidence (resolve_withfrom.py)

Every token on all five IBA rows, resolved through UniProt (with primaryAccession asserted
equal to the requested accession, size=10 so an ambiguous xref is reported rather than
silently reduced, and reviewed status tested with entryType.startswith("UniProtKB reviewed")),
then queried in QuickGO for its own evidence on the propagated term.

Per-row donor counts, exactly as GOA gives them:

So the "these donors only carry the same family-level inference" objection is false on every
row — it is testable here, and it fails, which is the expected outcome given that IBA
WITH/FROM lists experimentally-annotated members by construction.

Which term each donor actually holds (the ACRV1 question — not merely whether it holds one):

RETRACTED, and corrected here rather than deleted. An earlier draft of §1 stated that the
GO:0006355 donors disagree in sign — reading GO:0032786 as negative-branch by proximity to
GO:0032785. §3 refutes it: GO:0032786 is positive regulation of transcription elongation,
a descendant of GO:0045893, so every signed donor on that row points the same way, and
GO:0045893 is itself a descendant of GO:0006355, i.e. a positive child was available and
unused. The AEBP2 donor-disagreement test therefore does not apply.

What keeps the row at the unsigned parent is the recipient, not the donors: AFF3's own output
runs both ways — it represses XIST from the silent allele in HEK293T and IMR-90, while with
ZFP281 it establishes a permissive chromatin state at the Meg3 enhancer and its over-expression
raises 84% of the transcripts it changes in mouse cortical cells. A positive-only term would be
false for the repressive half. The specific negative instance is proposed as a separate
GO:0045892 row instead of by refining this one.

The GO:0050877 donors, by contrast, agree on GO:0007611 and the row sits two levels
above it. That is a granularity mismatch relative to the donors — but the specific term is not
supported for AFF3 itself, so the row is kept general rather than refined.

AFF3's own human nervous-system evidence then splits across the branch boundary, and §3 measures
where the boundary falls. GO:0050890 cognition is under GO:0050877, so the intellectual
disability, seizures and the GCC-expansion education association are inside the term and do
corroborate the row; GO:0021795 and GO:0001764 are not, so the cortical-migration evidence
is outside it and belongs on the separately proposed row. Offering both halves together as the
row's grounding — which an earlier draft did — conflates on-branch corroboration with off-branch
evidence, and that is what review round 1 caught.

2. The byte-identical WITH/FROM that means three different things

GO:0006354 carries PANTHER:PTN000829417|UniProtKB:P51825 on AFF1, AFF4 and AFF3 —
identical bytes in all three GOA records.

3. Ancestry claims, fetched not assumed (term_relations.py)

22 claims, all verified against QuickGO with relations=is_a,part_of only (so regulates
edges cannot be mistaken for subsumption). The script exits non-zero if any claim is wrong.

This guard has now caught two of my own claims, which is the argument for having it.

1. The GO:0030674/GO:0005515 claim was written the wrong way round on the first pass.
   It is retained in the script with the corrected expectation and a comment recording the error.
2. The GO:0032786 rows were added after the review was written, to check the §1
   sign-disagreement premise. They refuted it: GO:0032786 is positive, not negative, so the
   donors agree and the original argument for keeping GO:0006355 unsigned was false. The
   verdict survived on a different and better ground (the recipient's own mixed output), but the
   reason had to be rewritten. A claim I had already shipped, corrected by a check written
   afterwards.
3. The last four rows were added in review round 1. Three of them settle a tension the reviewer
   found in the GO:0050877 reason — the row was grounded in AFF3's human genetics, while the
   review argues elsewhere that AFF3's developmental nervous-system role is off-branch. The
   reviewer was half right: GO:0050890 cognition is under GO:0050877, so the
   intellectual-disability and language/education phenotypes are legitimate on-branch grounding,
   whereas GO:0021795 is not, so the migration evidence is the off-branch half. The reason
   now separates the two instead of lumping them, and rests primarily on the donor IMPs. The
   fourth closes the second leg of the GO:0003711/GO:0003712 sibling claim, which the reason
   asserted while only one leg was checked.

4. Reference-projection test (reference_projection.py)

Distinct gene products per reference (entities, not annotations; pagination asserted against
len(results), never against a page-size constant).

The finding. PMID:22547686 is the paper that isolated the AFF3-containing SEC-L3, and
its only annotation in all of GOA is on AFF1, the paralogue it contrasts SEC-L3 against.
AFF3 receives the term back as an IBA pointing at AFF1. This is the second instance of the
shape in this family: AFF4's review found PMID:20159561, titled for AFF4, likewise produced
one annotation and it was on AFF1.

The PMID:18616733 result is worth recording as a clean negative: the split is correct
curation, IEP for the mouse embryo in-situ and IMP for the human patient deletion, assigned
to the right species each.

5. Automated-route provenance

InterPro2GO — each of AFF3's three signatures, mapping fetched individually:

This reproduces AFF4's committed measurement exactly. No molecular-function term is produced
by any signature, so the fold-to-activity hypothesis does not confirm on AFF3 — reported as a
non-confirmation, not manufactured into a GO action.

ARBA — ARBA00026330, named in the GO_REF:0000120 row's WITH/FROM, has 1309 condition
sets; exactly one reaches AFF3:

IPR007797 AND IPR043640 AND taxon Eukaryota  ->  GO:0005634

The combinatorial reference is not three independent witnesses. GO_REF:0000120's tokens
are ARBA:ARBA00026330 | InterPro:IPR043640 | UniProtKB-SubCell:SL-0191. The ARBA rule's own
condition set is IPR007797 + IPR043640, so IPR043640 is counted twice; and SL-0191
derives from UniProt's own SUBCELLULAR LOCATION: Nucleus. line, which carries no evidence
tag. One signature, counted twice, plus UniProt citing itself. The row is still correct — AFF3
has its own nucleus IDA — but the apparent triple corroboration is illusory.

6. UniProt carries two GO terms GOA does not — and they resolve oppositely

QuickGO returns exactly 11 annotations for UniProtKB:P51826, matching the 11 rows of
AFF3-goa.tsv. UniProt's DR GO lines additionally carry two Ensembl-Compara projections:

GO:0003700 requires binding "a specific double-stranded genomic DNA sequence (sometimes
referred to as a motif) within a cis-regulatory region", and its usage comment warns against
this exact case. AFF3 is recruited by ZFP281/ZFP57 and its own DNA binding is non-specific.
The mouse source rows (PMID:25162227, IDA + IMP) rest on ChIP-qPCR of over-expressed HA-tagged
Laf4 at a single promoter, framed by the authors as testing a "potential direct transcriptional
regulator".

Correction to a sibling review's premise. AFF4's merged review states that AFF3's only
experimental molecular function is DNA-binding transcription factor activity. Read from AFF3's
own record, human AFF3 has no experimental molecular-function annotation at all — its only
MF row is the GO:0003712 IBA. The GO:0003700 IDA is on mouse Aff3.

A related stale projection, found while querying the donor: mouse Aff3 holds
GO:0016604 nuclear body and GO:0005829 cytosol by ISO GO_REF:0000119 from human
P51826, while human AFF3 currently carries neither.

7. Corrections check (corrections_check.py)

31 PMIDs cited anywhere in the review, the notes or the affinage record, checked by two routes:
PublicationType on the article, and CommentsCorrections/RefType on the article's own record
(the second catches Errata and Publisher Corrections that a publication-type search cannot see).

1 of 31 flagged; no retractions and no expressions of concern.

PMID:20444755 carries an ErratumIn reference to Ann Rheum Dis. 2011 Aug;70(8):1519 whose
PubMed id is null. Not resolvable by the two fallback routes either: Crossref returns empty
relation, update-to and updated-by for 10.1136/ard.2009.118406, and Europe PMC returns 0
hits both for that journal/year/page and for a citing Published Erratum. Its scope is therefore
unestablished — and the REMOVE verdict on that row deliberately does not rest on it.

8. Checks that came back negative, recorded so the next reviewer knows they were run

• Complex-projection test (a reference annotating a complex plus every subunit with
  identical evidence): none of the four references annotates more than two entities. Negative.
• Logical-opposite citation cross-product (a term and its negation sharing a reference set):
  AFF3's 11 terms contain no positive/negative regulation pair. Trivially negative.
• Per-partner GO:0005515 adjudication: AFF3 has no GO:0005515 rows in GOA, so there
  is nothing to adjudicate. IntAct nonetheless returns 14 records (all
  retrieved) over 7 distinct partners. Computed per-partner counts, from
  intact_partners.py / intact_partners.json:

The two SEC modules AFF3 bridges, CDK9 and MLLT1/ENL, are both present and GOA has curated
neither - an under-curation datum, not an over-annotation one. The scale of the gap,
measured across the family in one QuickGO call so a zero cannot be a rejected query:

The two paralogues are the positive controls: the endpoint works and the term is alive for
them in the same request, so AFF3's zero is a real absence. It is a curation asymmetry rather
than a biological one — AFF3's CDK9 contact is the most replicated of the three in IntAct. The first version of this
section was hand-counted and said "5 records across 4 distinct publications and 4 distinct
methods with MI 0.73" for CDK9. All four numbers were wrong. That is why the counts are now
derived from a committed script and quoted from its output table rather than written in prose.
- GOA-stub under-seeding (the ADAMTSL5 failure mode): the GOA TSV has 11 data rows and the
fetch-gene stub seeded 11 entries. They reconcile exactly; no GO:0005515 or
same-term/different-assigner rows were collapsed.
- Fold-to-activity propagation: not confirmed (§5). Fourth consecutive non-confirmation of
that lead in this campaign.