ADGRA2 bioinformatics — index

Hand-written index only. It deliberately does not restate any number: the numbers live in
the JSON outputs each script writes, and the interpretation lives in ../ADGRA2-notes.md.
Duplicating figures here would create a fourth surface to drift, which is the failure this
campaign has hit most often.

Every script re-runs from the repo with no arguments beyond those shown, and each writes its
JSON next to itself.

What each one found

Qualitative only — the figures live in the JSON, so there is no second surface to drift.

• node_reach.py — negative for a defect. The node whose human reach is exactly ADGRA2 gave
  it exactly ADGRA2's characterised biology; the generic terms sit at the heterogeneous family node.
  Specific-at-ortholog, generic-at-family is the correct direction. A null result is a finding: it
  tells the next reviewer the check was run, not skipped.
• resolve_partners.py — every GO:0005515 partner is a reviewed Swiss-Prot entry at
  canonical length carrying PDZ domains, from two independent PDZ-motif assays against one
  C-terminal motif. No TrEMBL or ORFeome substitution. This is what converts bare protein binding
  into GO:0030165 PDZ domain binding rather than into a list of biological interactors.
• interpro_signatures.py — the entry covering ADGRA2's transmembrane bundle also covers
  Frizzled proteins, and interpro2go deliberately maps it only to generic terms. The
  GPCR-activity claim comes instead from the family-2 extracellular hormone-receptor entries,
  i.e. a peptide-hormone ligand-recognition signature, on a receptor with no ligand.
• projection_test.py — the GDB TAS block is a projection: uniform evidence and assigner
  across every entity, one generic triple, no perturbed gene anywhere because the reference runs no
  functional assay. The per-entity matrix also settles the pseudogene sub-claim in the direction the
  review needs — the entities missing the molecular-function term are protein-coding, so both
  pseudogenes are among the recipients rather than among the omissions. This is the load-bearing
  evidence for the two REMOVE verdicts.
• check_corrections.py — no retraction and no expression of concern on any cited PMID; one
  Author Correction, which nothing in the review depends on.
• check_coverage.py / check_action_prose.py — invariant checks rather than analyses; they
  found real defects in the review itself (a collapsed GOA row set, and an over-annotation sentence
  sitting on a REMOVE row) and now prevent both from recurring.

Guards, and how to break-test them

check_corrections.py, check_coverage.py, interpro_signatures.py, projection_test.py
and check_action_prose.py all take --self-test, and each exercises its guard
in both directions — damage must be detected, and the clean file must pass. A self-test that
only proves failure detection cannot tell you the happy path works, and an agreement check that
fails on perfect agreement is a defect this campaign has actually seen.

node_reach.py and resolve_partners.py carry inline assertions rather than a self-test:

• node_reach.py asserts numberOfHits == len(results) after paging. This is not decoration.
  QuickGO caps this endpoint at 100 results per page and clamps rather than erroring on a
  larger limit, so a single limit=200 request on PTN001738137 returns 200 of 348 rows and
  looks complete. Under that truncation human ADGRA1 and ADGRA3 are absent from the result, which
  reads exactly like the known "family node misses the human paralogs" defect. The assertion is
  what turned that false positive back into a non-finding.
• resolve_partners.py builds its partner list from the GOA TSV rather than by hand, asserts the
  resolved count matches the parsed count, fails loudly on an accession that returns no name
  (a deleted UniProt entry is otherwise indistinguishable from an unannotated one), and tests
  Swiss-Prot status with entryType.startswith("UniProtKB reviewed") — because "reviewed" in entryType also matches "unreviewed" and silently promotes every TrEMBL entry to reviewed.