ADIPOQ (Q15848) — computed evidence for the GO annotation review

All numbers below are produced by analyze_adipoq.py and stored in
results.json. Re-run with:

uv run python analyze_adipoq.py            # full run (needs network)
uv run python analyze_adipoq.py --self-test  # break-test the guards

The script fails loudly on a missing input and never returns a partial
section. Every paginated query asserts numberOfHits == len(results); where a
reference is too large to paginate honestly the entity count is reported as
unavailable rather than estimated from one page.

A. The fetch-gene stub under-seeds this gene by 58 rows

There are no duplicate lines in the TSV — the 58-row gap is entirely the known
GOAValidator.seed_missing_annotations behaviour, which keys entries on
(GO id, evidence, reference, negated, qualifier) and omits WITH/FROM.
On ADIPOQ that collapses 65 IPI rows (one per interaction partner) into a
handful of stubs. The review restores one entry per GOA row; the guard in
analyze_adipoq.py asserts review_yaml_entries == goa_rows_distinct and
fails otherwise.

B. Reference-projection test — two bulk imports found

For every TAS / NAS / RCA / HDA / IPI reference, how many distinct
entities does that reference annotate in all of GOA?

Reading.

• GO:0140149 from PMID:36399478 is a database import. MatrisomeDB 2.0 is
  a database-update paper; one such reference giving 272 entities one
  identical cellular-component term is an import, not 272 author statements.
• GO:0005201 from PMID:28675934 is a bulk RCA block over 41 entities,
  alongside GO:0030020/GO:0030021/GO:0030023 (the collagen
  tensile-strength terms) over 31/8/4 more. The pipeline converted
  matrisome-proteomics detection into ECM structural roles.
• The GO:0005576 HDA row (PMID:27068509, 103 entities) is the same shape but
  the conclusion is independently true and experimentally established for
  adiponectin, so nothing turns on it.
• PMID:10095105 is a non-confirmation of the projection hypothesis: it has
  exactly one annotation in all of GOA — ADIPOQ's own GO:0006091 TAS.
  It is not an import; it is a single idiosyncratic annotation, and it must be
  judged on whether the paper supports the term (it is a gene-structure paper:
  "Organization of the gene for gelatin-binding protein (GBP28)").

C. The GO:0005515 set is one Y2H screen, counted five ways

The two-hybrid total is logged under five separate sub-method names —
two hybrid array (66), validated two hybrid (65),
two hybrid prey pooling approach (57), two hybrid pooling (4),
two hybrid bait and prey pooling approach (4). These are sub-methods of one
pipeline, which is why UniProt's NbExp=3 appears on nearly every partner.
There is no orthogonal biophysical assay anywhere in the set: no SPR, no
ITC, no co-IP of endogenous protein.

Topology. Adiponectin has a cleaved signal peptide (SIGNAL 1..18) and its
mature chain is secreted; Y2H requires both partners to reconstitute a
transcription factor in the yeast nucleus, which the native protein never
enters. Six partners are additionally annotated to cytosolic/nuclear/
mitochondrial compartments only — SGTA, COQ9, HTT, FASN, MRM1, TRIM35 — and
so could not meet secreted adiponectin even if the assay were physiological.
This is a supporting observation, not the main argument: the primary objection
is the assay, which places every partner outside its native compartment.

Negative results from this check, reported explicitly. All 61 partners
resolve to reviewed Swiss-Prot entries with canonical lengths — no TrEMBL
clones and no partial ORFeome constructs of the sort found on ACRV1. Thirteen
are cited as specific isoform accessions, which is normal for ORFeome-based
screens and is not itself a defect.

Partners that are not screen hits. PMID:12070119 (PDGFB, P01127) and
PMID:16622416 (ADIPOR1, Q96A54) are directed, hypothesis-led experiments,
and ADIPOR1 carries the reciprocal GO:0005515 IPI back to Q15848. They
are judged separately from the 59 screen partners.

D. All four IBA rows are self-referential and every donor is experimentally grounded

21 of 21 donor tokens resolved; 21 of 21 carry their own experimental
evidence. Every row's WITH/FROM includes UniProtKB:Q15848 itself — these
are self-referential IBAs, which record a PAINT curator judging the
function core, and are valid by construction.

MGI: and RGD: tokens are rejected by QuickGO's geneProductId with HTTP
400, so they are resolved through UniProt's xref:mgi- / xref:rgd- indexes
(bare numeric id — a query containing MGI's inner colon returns 400) and the
fallback route is recorded per token in results.json. MGI:MGI:106675
resolves to Q60994 ADIPO_MOUSE, the true mouse ortholog.

PAINT node placement is correct here — hypothesis not confirmed

The campaign's node-placement hypothesis (a term attached to the wrong node)
was tested in both directions and did not confirm:

• PTN008559544 reaches exactly one human gene, ADIPOQ — verified against
  all 82 human GO:0005179 IBA rows, all 3 GO:0045599 rows and the single
  human GO:0010642 IBA row, each query fully paginated. An ortholog-specific
  node carrying ortholog-specific terms is correct placement.
• PTN008355511 reaches ADIPOQ, C1QA, C1QB, C1QC, C1QTNF2, C1QTNF5,
  C1QTNF7, C1QTNF9 and C1QTNF9B and gives them GO:0005576 extracellular region. For a clade mixing complement C1q chains with the CTRP adipokines,
  the generic extracellular term is the LCA — refining it would mean picking
  one donor's compartment over another's.

Neither a misplaced term nor a mis-clustered member was found.

F. Ontology relations the review's prose asserts, fetched and checked

Regulation is not subsumption in GO, and — less obviously — the activity
branch is not under the binding branch. Both traps were live in this
review's first draft. Every relation the prose depends on is now fetched,
recorded in results.json under term_relations, and asserted:

The correction this check was written for. The first draft of the
GO:0005102 rows argued that the term was merely the coarser grain of
GO:0005179, "which is a descendant of GO:0048018 receptor ligand activity,
which is a descendant of this term". The last clause is false. GO:0048018
sits under GO:0140677 molecular function activator activity →
GO:0098772 molecular function regulator activity, in the activity branch;
GO:0005102 is a binding term under GO:0005515. Neither subsumes the other,
so the two rows are not redundant — they state different things (that
adiponectin physically engages a receptor, and that engaging it activates the
receptor), which strengthens the ACCEPT rather than weakening it.

Queries are is_a,part_of ancestor closures from QuickGO. The guard fails if
any observed membership disagrees with what the prose claims, and fails loudly
rather than vacuously if the assertion list is empty.

E. The two cold-thermogenesis terms are a 2 x 2 citation cross-product

GO:0120162 (positive regulation of cold-induced thermogenesis) and
GO:0120163 (negative regulation of cold-induced thermogenesis) are
logical opposites. GOA cites both terms to both of the same two references:

All four rows are ISS from UniProtKB:Q60994, all assigned by YuBioLab.
The detector reports is_full_cross_product: true.

This is a defect independent of what the papers say — the same evidence
cannot support a proposition and its negation. Reading the two abstracts
resolves which pairing is which:

• PMID:24531262 (Diabetologia 2014) — "Adiponectin reduces thermogenesis by
  inhibiting brown adipose tissue activation in mice"; Adipoq-/- mice have
  higher core body temperature and more UCP1. Supports the negative
  term only.
• PMID:26166748 (Cell Metab 2015) — "Adiponectin Enhances Cold-Induced
  Browning of Subcutaneous Adipose Tissue"; the thermogenic program is
  impaired in adiponectin-knockout mice. Supports the positive term
  only.

So two of the four rows are mis-paired with their reference
(GO:0120162/PMID:24531262 and GO:0120163/PMID:26166748), while the
remaining two are each correct. The underlying biological disagreement between
the two labs is genuine and unresolved; the annotation defect is separate from
it and is fixable without adjudicating the science.

Summary of what each check decided

Limitations

• This check-set matches computable structure (counts, ids, compartments). It
  cannot tell whether a well-formed annotation is biologically apt; those
  judgements are argued in ADIPOQ-ai-review.yaml and ADIPOQ-notes.md.
• The projection test is unavailable for the three largest interactome
  references (85343, 20010 and 3269 annotations). Their screen-artifact status
  rests on the IntAct method census in section C, not on an entity count.
• Section C's compartment call uses UniProt SUBCELLULAR LOCATION plus signal-
  peptide/transmembrane features. For multipass membrane partners it does not
  distinguish which loop faces which side, so the six-partner
  "cytosol-only" figure is a floor, not a total.