AP3S2 (P59780) — bioinformatics support for the annotation review

Three questions came out of the GOA record for human AP-3 complex subunit
sigma-2 (sigma-3B) that can be settled from sequence and database evidence
rather than from assertion. Each script fetches live from the UniProt, GO and
QuickGO REST APIs (responses cached under cache/, which is disposable) and
from the committed PANTHER PAINT slice
interpro/panther/PTHR11753/PTHR11753-paint.tsv. Nothing below is hardcoded;
delete cache/ and re-run to regenerate every number.

uv run python sigma_dileucine_pocket.py   # -> sigma_dileucine_pocket.tsv
uv run python resolve_withfrom.py         # -> withfrom_resolved.tsv
uv run python arba_funfam_specificity.py  # -> arba_funfam_specificity.tsv
uv run python check_goa_reconciliation.py # GOA tsv <-> review YAML

1. Does sigma-3B retain the dileucine-signal pocket? (sigma_dileucine_pocket.py)

Question. Acidic dileucine sorting signals [DE]XXXL[LI] bind no single AP
subunit; they bind a composite site across the large-subunit/sigma-subunit
interface — gamma1-sigma1 in AP-1, alpha-sigma2 in AP-2, delta-sigma3 in AP-3
(PMID:14691137, PMID:21097499). Mattera et al. (PMID:21097499) located the AP-3
site by substitution: "the loss of signal binding by the σ2 V88D or L103S
substitutions and the homologous σ1A V88D and I103S and σ3A V94D and L109S
substitutions". Every solved AP-3 structure to date is built on sigma-3A:
the 2024 cryo-EM core (PMID:39705307) and the 2026 AP3:ARF1 coat
(PMID:42139345), whose methods state the construct as "AP3S1(1–193) (H.
sapiens)". Whether the pocket also exists on sigma-3B is therefore a paralogy
question, and a falsifiable one.

Method. Canonical sequences for the four human sigma subunits used in
Mattera's substitution series, plus the mouse Ap3s2 ortholog named as donor by
the human ISS/IEA rows, are fetched live from UniProt. Each anchor residue is
first asserted against its own record, so a UniProt renumbering fails loudly
instead of silently reporting the wrong residue. AP3S2 is then globally aligned
to each comparator (Biopython PairwiseAligner, BLOSUM62, gap open −11 /
extend −1) and the AP3S2 position aligned to each anchor is read out.

Result.

Three things fall out.

• AP3S2 retains both pocket residues at the same native positions as AP3S1,
  Val94 and Leu109. The two paralogs differ at 31 of 193 positions overall
  (83.9 % identity), but only 4 of those 31 fall in the 46-residue window
  spanning the two anchors plus 15 residues either side — V113M, N118Y, A121Q
  and M123V, all downstream of Leu109 and none at an anchor. The structural work
  done on sigma-3A therefore transfers to sigma-3B at this site, and no
  loss-of-pocket argument is available against AP3S2.
• The alignment independently reproduces the equivalence Mattera asserted
  across complexes: sigma-2 V88/L103 and sigma-1A V88/I103 both map onto AP3S2
  V94/L109. The one SUBSTITUTED call is Ile→Leu at the second position, which
  is the difference Mattera's own text reports between sigma-1A and the
  sigma-2/sigma-3 pair, not a divergence of AP3S2.
• Human and mouse AP3S2 are byte-for-byte identical proteins — both 193 aa,
  100 % identity, no gaps. This is directly relevant to grading the ISS rows
  whose donor is UniProtKB:Q8BSZ2: the donor is not merely the true ortholog,
  it is the same sequence.

Limits. This is a sequence argument about a binding site, not a measurement
of binding. It shows the pocket is present and unchanged; it does not show that
a delta-sigma3B hemicomplex binds a given cargo with the same affinity as
delta-sigma3A. The direct evidence for that is experimental and is discussed in
the notes (PMID:14691137 tested delta-sigma3B explicitly).

2. Who are the IBA donors and the PAINT seeds? (resolve_withfrom.py)

Question. The single IBA row (GO:0016192 vesicle-mediated transport)
carries eleven gene-product donors plus the ancestral node
PANTHER:PTN000204281. Reviewing it means knowing what those donors are, and
which complex each one belongs to.

Method. Every token in the GOA WITH/FROM column and every seed in the
committed PAINT slice is resolved through the UniProt cross-reference search
(query=xref:<db>-<id>, size=5, so multi-hits are reported rather than
silently collapsed), preferring the reviewed entry. WormBase gene ids are not in
UniProt's xref index; those fall back to the GO API bioentity record and then
to a gene-name lookup, rather than being written off as unresolvable.

Result. 30 tokens; all resolved (one via the GO API fallback). The slice
holds two IBD assertions on the same node:

The ten seeds of the GO:0016192 IBD resolve to:

The IBA row carries one further donor not in this seed list,
FB:FBgn0043012 (D. melanogaster AP-2sigma, Q9VDC3); that gene is a seed of
the same node's GO:0043231 IBD, whose slice line is three months older. The
discrepancy is a release-timing artefact between the GOA row and the current
PAINT slice, not a donor the tree does not contain.

Interpretation for the review. Two of the ten seeds are genuine AP-3 sigma
orthologs, so the node is not an AP-1/AP-2-only inference being stretched to
AP-3. The remaining eight are AP-1 and AP-2 sigma subunits, which act on
different itineraries (TGN/endosome, plasma membrane) from AP-3's
(endosome/lysosome). Donors that disagree in destination are exactly the case in
which the parent term is the correct least common ancestor: GO:0016192 vesicle-mediated transport is general because the family is, not because the
curator under-called it.

3. Is the ARBA rule behind the GO:0030123 IEA AP-3-specific? (arba_funfam_specificity.py)

Question. The GO:0030123 AP-3 adaptor complex / IEA / GO_REF:0000120 row
cites ARBA:ARBA00033921|InterPro:IPR027155. FunFam-based rules are named for
whole families, so a rule of that shape can assign an AP-3-specific complex term
to every sigma subunit in the fold. Whether this one does is empirical.

Method. The rule is fetched live from https://rest.uniprot.org/arba/ and
its condition set printed; the FunFam cross-reference is then read off every
reviewed human AP-complex sigma subunit (the complete set, not a sample).

Result. ARBA00033921 (created 2022-04-29, modified 2025-03-21) fires on
FunFam id = 3.30.450.60:FF:000001 and taxon = Primates, and asserts
GO:0030123.

The panel is the complete set of reviewed human AP-complex sigma subunits. The
first seven are also the complete human membership of PANTHER family PTHR11753,
checked against interpro/panther/PTHR11753/PTHR11753-entries.csv; AP5S1 is the
one that sits outside the family, and is included so the test is not scoped to
the family the rule is about.

The suspicion is not borne out. Within the human sigma panel the FunFam
partitions cleanly by complex, and exactly the two AP-3 sigma subunits match.
Zero non-AP-3 subunits would be called AP-3 by this rule. The IEA row is
therefore well-founded on both halves of its WITH/FROM, and no
PROPAGATION_BAD call is warranted against it.

Limits. The panel is human. UniProt's search index does not expose FunFam as
a queryable cross-reference, so the rule's precision across all primates was not
measured; what is measured is that the discriminating signature exists and
separates AP-3 from AP-1, AP-2 and AP-4 in the species being annotated.

4. GOA ↔ review reconciliation (check_goa_reconciliation.py)

Mechanical check, run before every commit: every row of AP3S2-goa.tsv maps to
exactly one existing_annotations entry on (GO id, evidence code, reference,
qualifier, normalised WITH/FROM); every non-NEW entry maps back to a GOA
row; supporting_entities is the |-split WITH/FROM verbatim; and every
IBA/ISS/ISO/IEA/IC row that carries a WITH/FROM carries a
propagation_review. 25 GOA rows; 16 of them require a propagation_review.
Exit status is non-zero if any check fails.