AP3D1 (O14617) — bioinformatics support for the annotation review

Two questions from the GOA record for human AP-3 delta can be settled from
sequence and from the records the pipeline already commits, rather than from
assertion. Both scripts fetch live (UniProt REST, cached under the disposable
cache/) and read the committed PANTHER slice
interpro/panther/PTHR22781/PTHR22781-entries.csv. Nothing below is hardcoded;
delete cache/ and re-run to regenerate every number.

uv run python delta_interfaces.py      # -> delta_interfaces.tsv   (needs mafft on PATH)
uv run python goa_reconcile.py         # -> pass/fail, counts

1. Are the delta ARF1 and VAMP7 interfaces family features? (delta_interfaces.py)

Question. Three interfaces on AP-3 delta have been mapped at residue
resolution, and all three were mapped on human AP3D1 itself, so their
positions need no cross-species inference:

PMID:39705307 independently localises the primary ARF1 site to delta by
hemicomplex pulldown, and its cryo-EM depositions 9C58/9C59/9C5B/9C5C map to
O14617 1–617. The review needs to know two things these papers do not state:
whether human really carries those residues, and whether the interfaces are
delta-family features or generic adaptin features.

Method. One MAFFT --auto alignment over a panel that is derived, not
hand-picked: every reviewed member of PTHR22781 in the committed InterPro
slice (9 proteins, all PTHR22781:SF12, human through Arabidopsis), plus the
four human large subunits of the other adaptor complexes as out-of-family
controls — AP3B1 (PTHR11134), AP1G1, AP2A1 and AP4E1 (all PTHR22780). The
column each human position occupies is read off, and each panel member's own
native position is reported alongside its residue.

The script asserts len(O14617) == 1153 and asserts each of the eleven claimed
residues before aligning, so a changed sequence or a mistyped position fails
loudly instead of producing a plausible table.

Results (delta_interfaces.tsv; identity to human at the site positions):

Verified, first. All eleven residues are present in O14617 at exactly the
published positions (the pre-alignment assertions pass). The papers' numbering
is human numbering.

ARF1 site 1 is an ancient delta feature. 3/3 from budding yeast to human
inside the family; 1/3 in every out-of-family control. Consistent with
PMID:39705307's finding that the delta site is structurally the analogue of the
AP-1 gamma / COPI site while the primary site differs between complexes.

ARF1 site 2 is a basic patch, mammal-tight and deeper-taxon-loose. 4/4 in
mouse and bovine, 1–2/4 outside vertebrates. This is weaker than "universally
conserved" and matches the paper's own hedged wording ("highly conserved among
multicellular eukaryotes"). AP4E1 scoring 3/4 is the expected result rather than
a contradiction: the basic ARF1-contact patch is a general adaptin large-subunit
feature, not delta-private — what is delta-private, per PMID:39705307, is that
this is AP-3's primary rather than secondary site.

The VAMP7 hinge is metazoan-enriched and delta-private. 0/4 or 1/4 in all
four out-of-family human controls; 1/4 or less in the three fungi; 3–4/4 across
metazoa and Dictyostelium. Two independent signals agree, because the hinge
sits inside InterPro IPR010474, whose name is literally "AP-3 complex subunit
delta domain, metazoa".

Caveat. The hinge lies between UniProt's two MobiDB-lite disordered regions
(629–696 and 726–920), so column assignment there is less reliable than in the
HEAT trunk, and the fungal/plant scores should be read as "not recoverable from
this alignment" rather than as demonstrated loss. The mouse and bovine results
do not depend on the alignment being right in a hard region: both proteins come
out with native positions identical to human's for all eleven residues
(native_position column), which is what the review's residue_claims use as
anchor coordinates.

Use in the review. The Bilateria IBD node PTN000513028 is seeded by mouse
Ap3d1 (plus rat, for GO:0016182). Recording that human retains every mouse
residue at these interfaces is the checkable form of "there is no
target-specific evidence of divergence that would undercut the transfer" —
encoded as RETAINED residue_claims with anchor UniProtKB:O54774, target
UniProtKB:O14617, method MSA.

2. Does the review YAML still match the GOA file? (goa_reconcile.py)

Question. existing_annotations is seeded deterministically from
AP3D1-goa.tsv and then hand-edited. Nothing in just validate proves the
edited file still corresponds 1:1 to the GOA rows, or that a
propagation_review.source_entities entry names a source the row actually
carries.

Method. Both sides are reduced to a multiset keyed on
(GO id, evidence code, reference, normalized WITH/FROM set) and compared. The
script additionally checks that every IBA row, and every ISS/ISO/IEA/IC row with
supporting_entities, has a propagation_review; that every
source_entities[].source_id appears in that row's own supporting_entities
(the defect that drift produces when these lists are hand-typed); and that every
reference cited anywhere is declared in references:.

It also runs a strict verbatim check on every supporting_text: a literal
substring test after whitespace normalisation only. This is deliberately stricter
than the repo/CI reference validator and than checkquotes.py, which both
normalise before matching — and that normalisation silently accepts a quote that
transliterates a character the paper prints, e.g. delta-adaptin where
PMID:22521722 has δ-adaptin. Such a quote passes every existing check while no
longer being findable in the paper. Two quotes in this review were caught that
way and restored to the Greek character.

Result. 59 GOA rows ↔ 59 GOA-derived YAML entries, no key unmatched in
either direction. Evidence codes: IEA 29, IBA 10, NAS 8, IMP 3, HDA 2, TAS 2,
ISS 2, IPI 1, IDA 1, IC 1. 42 rows require a propagation_review (10 IBA plus
32 IEA/ISS/IC rows carrying supporting_entities). Every count quoted in
AP3D1-notes.md, in the review YAML and in the PR body comes from this script's
output rather than from a hand tally. 135 supporting_text quotes pass the strict
verbatim check.