AP3S1 (sigma3A) — review notes

UniProt Q92572, 193 aa, AP3S1_HUMAN. Accession asserted against the fetched record:
ID AP3S1_HUMAN Reviewed; 193 AA. / AC Q92572; — this is the
expected protein, not a merged accession returning something else. PANTHER family
PTHR11753 (ADAPTOR COMPLEXES SMALL SUBUNIT FAMILY), shared with AP1S1, AP1S2, AP1S3,
AP2S1, AP3S2 and AP4S1. HGNC:2013, 5q22.3-q23.1, previous symbol CLAPS3 (HGNC REST,
checked live).

GOA seeded 30 rows, and the evidence-code census (scripted off the tsv) is one IBA, one IDA,
one IPI, two ISS, fourteen IEA, eight NAS and three TAS; by assigner, eight ComplexPortal,
six UniProt, five InterPro, four Ensembl, three PINC, two GOC, one GO_Central and one
FlyBase. Everything mechanistic about this protein is in papers GOA has not used.

1. What the protein is

sigma3A is the small subunit of AP-3, the adaptor that sorts transmembrane cargo out of
tubular sorting endosomes towards lysosomes and lysosome-related organelles (LROs). The GO
definition of GO:0030123 names the subunit and both of its isoforms explicitly, which
settles the complex-membership rows without argument: "A heterotetrameric AP-type membrane
coat adaptor complex that consists of beta3, delta, mu3 and sigma3 subunits and is found
associated with endosomal membranes. AP-3 does not appear to associate with clathrin in all
organisms. In at least humans, the AP-3 complex can be heterogeneric due to the existence
of multiple subunit isoforms encoded by different genes (beta3A and beta3B, mu3A and mu3B,
and sigma3A and sigma3B)."

UniProt's FUNCTION is complex-level and reflects the 1997 literature:
CC -!- FUNCTION: Part of the AP-3 complex, an adaptor-related complex which is …
"It facilitates the budding of vesicles from the Golgi membrane and may be directly
involved in trafficking to lysosomes." The SUBUNIT line gives the composition and
partners: two large adaptins (AP3D1 plus AP3B1 or AP3B2), a medium adaptin (AP3M1 or
AP3M2), a small adaptin (AP3S1 or AP3S2), interaction with AGAP1, and association with
BLOC-1. Localisation is Golgi apparatus and cytoplasmic vesicle membrane, peripheral, on
the cytoplasmic side.

sigma3A and sigma3B were identified together in the paper that named AP-3
PMID:9118953,
and independently as subunits of the same complex in
PMID:9151686.

Where the complex acts has moved since 1997. Simpson's immunofluorescence put AP-3 "in the
Golgi region of the cell as well as with more peripheral structures"; the compartment that
survived is endosomal
PMID:15051738
and, in pigment cells,
PMID:16162817.
This matters for the Golgi to vacuole transport row in §6.

A stale locus in the citation trail. The AP-type membrane coat adaptor complex TAS
row cites PMID:8697810, whose title asserts "mapping to 12p 13.2 --> p13.1 of CLAPS3". That
FISH assignment is superseded: HGNC places AP3S1 at 5q22.3-q23.1, UniProt files it under
Proteomes; UP000005640; Chromosome 5 and CCDS4123.1. The paper's other content — the 193
aa ORF, the ubiquitous 1.35 kb transcript — is correct and is what the annotations rest on,
so this is a footnote, not a reason to doubt the rows.

2. The molecular function: sigma3A is half of the AP-3 dileucine cargo-signal site

This is the substantive finding, and it is entirely absent from GOA. The AP-1 case was
worked out in this repository's AP1S1 review; most of the
reasoning transfers, and §2.3 below records where it does not.

2.1 What transfers unchanged

Cargo signals come in two flavours and bind different parts of an AP complex. Tyrosine
YXXØ motifs bind the mu subunit. Acidic dileucine motifs bind no single subunit — they
bind the large-subunit/sigma hemicomplex:

• PMID:14691137
• PMID:21097499
• PMID:21097499

The "not any single subunit" clause is what the contributes_to qualifier exists for, and
it is why the core function below uses contributes_to_molecular_function.

2.2 The evidence is on sigma3A specifically, not on AP-3 generically

This is the distinction a subunit review is supposed to draw, and for AP-3 it is unusually
clean, because the 2011 mutagenesis paper mutated sigma3A itself:

• PMID:21097499
• PMID:21097499
• PMID:21097499

So for AP-3 the acidic-residue contact is genuinely shared between delta and sigma3A,
whereas in AP-1 it sits mainly on gamma1. Four distinct signals were tested against
delta-sigma3A and bound
PMID:21097499,
and the original demonstration was
PMID:14691137.

The specific cargo that GOA already records as an AP3S1 interactor — tyrosinase, the
protein binding IPI partner — is a dileucine cargo of exactly this site
PMID:16162817 and
PMID:16162817.
That is the functional content the bare GO:0005515 protein binding row throws away.

2.3 Where the AP-1 reasoning does not transfer

Three places:

1. The GO term to use. The AP1S1 review annotated the hemicomplex activity through
   GO:0035615 clathrin-cargo adaptor activity, noting that the definition — "Bringing
   together a cargo protein with clathrin, responsible for the formation of endocytic
   vesicles" — names the wrong compartment for AP-1. For AP-3 that mismatch is worse, not
   equal: AP-3 is not an endocytic coat at all, it acts on endosomal exit
   PMID:15051738, and AP-3's clathrin association is itself conditional (the GO:0030123
   definition says AP-3 "does not appear to associate with clathrin in all organisms", and
   a beta3A point mutant with a dead clathrin box gives full functional rescue
   PMID:11807095).
   So this review uses the parent, GO:0140312 cargo adaptor activity, instead. I checked
   that GO:0035615 is GO:0140312's only child via the QuickGO children endpoint, so
   the parent is not a lazy generalisation — it is the most specific term in that lineage
   that does not assert endocytosis.
2. The acidic-residue contact. In AP-1 it is mainly on the large subunit; in AP-3 both
   delta R26 and sigma3A R15 contribute (§2.2). The sigma subunit carries proportionally
   more of the AP-3 site than of the AP-1 site.
3. A structure of the assembled complex exists for AP-3 too, and it says something the
   AP-1 structure did not. In the human AP-3 cryo-EM structures the pocket is not empty
   and not occupied by cargo: PMID:39705307,
   with the authors proposing that "full activation requires ejection on the β3 tail,
   potentially by engagement of a dileucine cargo motif". The AP-1 structure
   (PDB 4P6Z) had a dileucine mimic bound in the sigma1 pocket; the AP-3 structure has an
   autoinhibitory occupant in the sigma3 pocket. Same site, opposite snapshot.

2.4 There is still no GO term for it

Checked independently rather than inherited from the AP1S1 review: QuickGO's ontology
search returns zero hits for dileucine; targeting sequence binding returns
GO:0089710 endocytic targeting sequence binding, whose definition requires "an essential
tyrosine (Y)" and internalization by clathrin-coated pits, i.e. the other motif class; and
the molecular-function children of GO:0140312 cargo adaptor activity are exactly one
term, GO:0035615. There is no dileucine counterpart. Hence a proposed_new_terms entry
and an ONTOLOGY knowledge gap, consistent with the one AP1S1 raised — the same term would
serve both genes and their large-subunit partners.

3. A sigma3A-specific regulatory mechanism GOA does not record

AP-3 membrane recruitment is controlled by an interaction that targets sigma3 directly:
PMID:15469849,
with the consequence that "the delta-ear inhibits the recruitment of AP-3 to membranes both
in vitro and in vivo and impairs the sorting of lysosomal membrane proteins". This is a
regulatory interaction on sigma3, not an activity of sigma3, so it does not become an
annotation here; it is recorded as a suggested question and as context for the structural
role in §4.

4. The structural role, and why it is worth a NEW row

GOA gives AP3S1 exactly one molecular-function term, GO:0035651 AP-3 adaptor complex binding, which §7 argues should go. The subunit's actual subunit-level job — holding the
complex together — is annotated nowhere.

The evidence is direct and in human protein:

• The human complex was reconstituted from its four subunits including sigma3A and solved
  by cryo-EM PMID:39705307.
  UniProt maps four of those entries to this protein — DR PDB; 9C59; EM; 4.30 A; S/s=1-193.
  and three others — so chain S of the solved AP-3 core is AP3S1, full length.
• sigma3 is one half of a stable hemicomplex with delta, and the pairing survives loss of
  the other half: PMID:39705307
  (the Arf1 site itself is on delta, not sigma3 — this is not a claim that sigma3A binds
  Arf1).
• Genetically, in mouse cells lacking one or the other large subunit, sigma3's partner
  dependence is explicit: PMID:11807095

Coded IDA on PMID:39705307, because the reconstitution and the structures are direct assays
of the human protein's contribution to the assembled complex; the mouse hemicomplex
genetics is cited as corroboration, not as the human evidence.

5. The IBA: one row, the pan-AP-sigma node, and the right LCA term

The single IBA is GO:0016192 vesicle-mediated transport, GO_REF:0000033, WITH/FROM
listing 11 gene-level donors plus PANTHER:PTN000204281.

The family PAINT slice is already in the repo
(interpro/panther/PTHR11753/PTHR11753-paint.tsv). PTN000204281 carries two IBD
assertions:

UniProtKB:Q92572 is not among the seeds of either IBD, so this is not a
self-referential IBA and there is no redundancy question either way. (GO:0043231 does not
appear as an IBA row in AP3S1's GOA, so there is nothing to review for it.)

Every donor resolved:

Eleven gene-level donors spanning AP-1, AP-2 and AP-3 sigma subunits across fungi,
nematode, fly, rodent and human. The node is the deep pan-AP-sigma ancestor, not an AP-3
node.

That is why GO:0016192 is the correct term and not a granularity failure. The tempting
review is "vesicle-mediated transport is vague; AP-3 does endosome-to-lysosome delivery, so
MODIFY". GRANULARITY_MISMATCH applies only when the donors agree and a more specific term
was available. Here they do not agree: the AP-1 sigmas work at the TGN/endosome interface,
the AP-2 sigmas in endocytosis at the plasma membrane, the AP-3 sigmas in lysosome/LRO
delivery. vesicle-mediated transport is precisely their least common ancestor, and a
compartment-specific term at this node would over-propagate onto the AP-1 and AP-2 members.
Verdict NO_FAILURE_CORE.

Two AP-3-clade donors carry their own experimental grounding for the fungal form of this
process, which is the strongest link from the node to AP3S1's own subfamily: yeast Aps3
GO:0006896 IMP PMID:9335339 (plus GO:0030123 IMP/IPI PMID:9250663), and C. albicans Aps3
GO:0006896 IMP PMID:20870878 (QuickGO, checked live).

One discrepancy worth recording, checked not assumed. The row's WITH/FROM carries
FB:FBgn0043012, but the GO:0016192 IBD seed list in the fetched PAINT slice does not —
it has 10 seeds, not 11. The slice's GO:0016192 row is dated 20260828 and the GOA row
20250902, so this is a snapshot difference rather than a defect; FBgn0043012 does appear
in the same node's GO:0043231 seed list. The same difference is visible in the AP1S1
review, independently.

Every human sigma paralog gets this identical row. Queried via QuickGO: AP1S1
(P61966), AP1S3 (Q96PC3), AP2S1 (P53680), AP3S1 (Q92572), AP3S2 (P59780) and AP4S1
(Q9Y587) all carry GO:0016192 IBA with a byte-identical 12-entry WITH/FROM. Note that
AP2S1 is both a donor and a recipient — for that gene the IBA is self-referential and
legitimately so; for AP3S1 it is not.

6. The InterPro and UniProt-vocabulary rows

Three InterPro signatures carry six GO mappings between them, producing the five
GO_REF:0000002 rows and contributing to the one GO_REF:0000120 row (counted off the tsv,
not estimated). IPR000804 (clathrin adaptor small-chain conserved site, matched here as
PROSITE PS00989 on the UniProt record) maps to GO:0006886, GO:0016192 and GO:0030117 — all
three true of any AP sigma. IPR016635 ('Adaptor protein complex, sigma subunit') maps to
GO:0015031 only. Those are all correct and appropriately general.

IPR027155 is the interesting one. It is the AP-3-sigma-specific entry (APS3), and it
maps to GO:0006896 Golgi to vacuole transport. That term is real and not fungus-only —
GO:0090160 Golgi to lysosome transport is its child and lysosome is_a vacuole, so nothing
excludes mammals a priori — but the donor compartment it names is the fungal ALP pathway,
where the seed evidence sits (yeast and C. albicans Aps3, both IMP, §5). Human AP-3's
demonstrated route starts at the endosome, not the Golgi: PMID:15051738 again, and
PMID:16162817.
So the row is compartment-shifted rather than wrong, and the review MODIFYs it to
GO:0008333 endosome to lysosome transport (verified: BP, not obsolete, is_a GO:0007041 /
GO:0016192).

The ARBA half of the GO:0030123 IEA was fetched and read
(https://rest.uniprot.org/arba/ARBA00033921): one condition set, FunFam id = 3.30.450.60:FF:000001 AND taxon = Primates, annotating GO:0030123. That looks loose —
a FunFam plus a clade — so I checked whether the FunFam actually separates AP-3 sigmas from
the rest of the family. It does: among the seven human sigma paralogs, only AP3S1 and
AP3S2 carry 3.30.450.60:FF:000001; AP1S1 and AP1S3 carry FF:000005, AP1S2 FF:000009,
AP2S1 FF:000004 and AP4S1 FF:000010 (UniProt xref_funfam, queried live). The rule is
discriminative. It does not separate sigma3A from sigma3B — but neither does anything else
(§8), and GO:0030123 is correct for both.

7. The one annotation that is actually wrong: AP-3 adaptor complex binding

GO:0035651, MF, IEA GO_REF:0000107, WITH/FROM UniProtKB:Q9DCR2|ensembl:ENSMUSP00000025357.

The term's definition is "Binding to an AP-3 adaptor complex", and AP3S1 is an AP-3
adaptor complex subunit. GO already expresses that relation, three times over, as
GO:0030123 part_of (IDA, NAS and IEA rows on this gene). Typing a constitutive subunit as
a ligand of its own complex is a role conflation, and it is the whole reason the
part_of relation exists.

This is not a one-off. A QuickGO query for GO:0035651 restricted to human returns 29
annotations over 28 entities, and all seven human AP-3 subunit genes are among them —
AP3B1, AP3B2, AP3D1, AP3M1, AP3M2, AP3S1, AP3S2 — every one of them by IEA. The entities
with experimental evidence for the term are the ones that genuinely bind AP-3 from outside:
PI4K2A (IDA), RAB32 and RAB38 (IPI). So the complex's own subunits have been swept in
alongside its binding partners.

The proximate source is the mouse Ap3s1 row, GO:0035651 IDA PMID:19010779 (QuickGO). That
paper is an AP-3 cross-linking/mass-spectrometry study in which AP-3 was the bait and
BLOC-1, BLOC-2, HOPS, clathrin and PI4KIIalpha were the associated proteins
PMID:19010779 —
i.e. the subunits are the complex, not its ligands. The cache is abstract-only for that
paper, so I am not second-guessing an assay I cannot see; the argument does not need the
assay. It is that no experiment can make a constitutive subunit a ligand of the complex it
constitutes, and the row being reviewed here is the human IEA, not the mouse IDA.
REMOVE, root_cause: TERM_SCOPING_PROBLEM, failure_modes: [ROLE_CONFLATION].

8. sigma3A versus sigma3B: inherited, and correctly so

The prompt asked whether GOA distinguishes the two sigma3 isoforms or whether AP3S1 simply
inherits. Answer, from a QuickGO diff of Q92572 (30 rows) against P59780 (25 rows):

• The complex-level rows are shared verbatim: the same GO:0030123 IDA (PMID:9118953),
  the same ComplexPortal NAS set (PMID:9151686, PMID:9545220, PMID:15537701, PMID:23247405),
  the same InterPro2GO rows, the same ARBA rule, and the byte-identical GO:0016192 IBA.
• The ortholog-transfer rows are paralog-correct, not confused: AP3S1's Compara/ISS
  donor is mouse Ap3s1 (Q9DCR2) and AP3S2's is mouse Ap3s2 (Q8BSZ2). I checked both
  accessions in UniProt rather than assuming.
• Scripted diff on (GO id, evidence code, reference): QuickGO returns 30 rows for Q92572 and
  25 for P59780, of which 24 are shared. Six rows are unique to AP3S1 -- the three PINC
  TAS rows (PMID:8697810 cloned CLAPS3 = sigma3A; PMID:9118953; PMID:9792713), the
  protein binding IPI to tyrosinase (PMID:16162817), a membrane Compara row and the
  presynapse inter-ontology row that follows from AP3S1's synaptic vesicle coating row.
  The first four come from sigma3A-specific papers; the last two are pipeline consequences,
  not new biology. One row is unique to AP3S2, a synaptic vesicle IEA transferred from rat
  (UniProtKB:A0A0G2K302).

That is not curation laziness. The biochemistry says the two are interchangeable at the
tested site: PMID:14691137
and PMID:14691137.
The bioinformatics in §9 adds the sequence-level counterpart: the two proteins differ at 31
of 193 positions but are identical at every pocket residue the literature names.

Worth flagging for the sibling AP3S2 review: the Nef result does separate the two large
subunits, but not the sigmas — only two of five Nef variants bound delta-sigma3A
PMID:14691137 —
and sigma3B was not tested against Nef at all.

9. Bioinformatics: the site is retained; retention is not the evidence

AP3S1-bioinformatics/ (uv run python sigma3_pocket.py) fetches ten sigma subunits live
from UniProt, asserts each length, and maps the structurally defined sigma2 pocket onto
each by pairwise alignment. Results in
file:human/AP3S1/AP3S1-bioinformatics/RESULTS.md. Four outcomes:

1. All five literature-named sigma3A positions verify against the live Q92572 sequence:
   R15, V94, D98, L107, L109.
2. The alignment independently reproduces both correspondences PMID:21097499 asserts
   (sigma2 88 -> sigma3A 94, sigma2 103 -> sigma3A 109), and the sigma1A pair as well. The
   mapping method is sound.
3. The negative control defeats the conservation argument. sigma4 (AP4S1) is in the
   same PANTHER family and its hemicomplex demonstrably does not bind these signals, yet it
   scores 4/5 identical at the pocket — the same as sigma1A, sigma1B and sigma1C, all
   binders. sigma3A scores 5/5, a better match to the sigma2 anchor than any AP-1 sigma
   (Leu, not Ile, at the sigma2 L103 position), but 5/5 versus 4/5 is not a separation when
   the non-binder is inside the 4/5 group. So "AP3S1 retains the dileucine pocket" is true
   and carries no weight by itself; the direct sigma3A mutagenesis does.
4. sigma3A vs sigma3B: 83.9% identical overall (31 differing positions), identical at all
   five pocket positions; and human AP3S1 vs mouse Ap3s1 is 100% identical over 193 aa,
   which is why the Compara and ISS rows in §10 are unusually safe transfers.

Recorded as residue_claims (RETAINED) on the molecular-function row, because that is the
checkable form — while the propagation reasoning for the IBA is explicitly not
residue-based and says so.

10. The Ensembl-Compara and ISS rows, and the perturbation behind them

Four rows come from GO_REF:0000107 with UniProtKB:Q9DCR2|ensembl:ENSMUSP00000025357,
and two more restate two of them as curator ISS (GO_REF:0000024). Q9DCR2 is mouse Ap3s1,
checked in UniProt — the 1:1 ortholog, not the sigma3B paralog, and sequence-identical to
the human protein (§9).

The mouse sources, from QuickGO:

The perturbation in PMID:21998198 is not an Ap3s1 mutant. The paper's genetics is the
mocha allele, Ap3d1mh/mh — a delta-subunit null that removes the whole complex — and
sigma3 appears as an assay readout rather than as the perturbed gene
PMID:21998198.
Annotating an obligate-complex subunit from a complex-null phenotype is standard GO
practice and I am not calling it a defect. It does determine the grade: the claim is
complex-level and neuronal-context, so the axonal/synaptic-transport rows and the
inter-ontology axon cytoplasm row that follows from them are KEEP_AS_NON_CORE, not
core, for a subunit of a ubiquitously expressed adaptor. root_cause: NO_FAILURE_NON_CORE.

GO:0016020 membrane is root-ish and uninformative — AP-3 is a peripheral membrane coat,
so it is true of the whole complex — and its mouse source is a CACAO IMP on a
BLOC-1/AP-3 interaction study PMID:16837549. KEEP_AS_NON_CORE.

11. The ComplexPortal NAS rows, sized by projection

Eight rows come from ComplexPortal as NAS. Rather than guess how gene-specific they are, I
counted entities per reference in QuickGO (paginated; entity counts, not annotation
counts):

So the melanosome-assembly, platelet-dense-granule and early-endosome rows are one
statement about AP-3 projected across every subunit, and they are graded
KEEP_AS_NON_CORE accordingly — real biology, cell-type-restricted, not evidence about
sigma3A. The underlying claims do hold for AP-3: pigment and platelet phenotypes are the
defining consequence of AP-3 loss
PMID:9697856
and PMID:10024875,
and there is a gene-level in-vivo perturbation of the ap3s1 orthologue itself
PMID:20713646.

GO:0035654 clathrin-coated vesicle cargo loading, AP-3-mediated is the exception among
the NAS rows: its definition is "Formation of a macromolecular complex between proteins of
the AP-3 adaptor complex and proteins and/or lipoproteins that are going to be transported
by a clathrin-coated vesicle", which is exactly the cargo-recognition step sigma3A performs
(§2). ACCEPT, core.

GO:0016183 synaptic vesicle coating is the one NAS row whose term is wrong. Its
definition is "The formation of clathrin coated pits in the presynaptic membrane endocytic
zone, triggered by the presence of high concentrations of synaptic vesicle components" —
plasma-membrane coated pits, which is AP-2's job, not AP-3's. AP-3's synaptic role is
budding vesicles from endosomes, and GO has a term for that: GO:0016182 synaptic vesicle budding from endosome, defined as "Budding of synaptic vesicles during the formation of
constitutive recycling vesicles from early endosomes". MODIFY. The sibling row
GO:0098793 presynapse, which the inter-ontology pipeline derives from GO:0016183, is kept
as non-core on its own merits: sigma3A is a subunit of the neuronal AP-3 variant as well as
the ubiquitous one (UniProt lists ComplexPortal; CPX-5055; Neuronal AP-3 Adaptor complex, sigma3a variant), and the neuronal complex is targeted to processes
PMID:15537701.

12. Insulin receptor signalling

GO:0008286, TAS, PMID:9792713. The paper found sigma3A in an expression-library screen
for IRS-1 binders in 3T3-L1 adipocytes and closed with a hypothesis: "These results are
consistent with the hypothesis that sigma3A serves as an IRS-1 receptor that may dictate
the subcellular localization and the signaling functions of IRS-1." No step of the insulin
receptor signalling cascade — the term's definition is "The series of molecular signals
generated as a consequence of the insulin receptor binding to insulin" — is shown to be
executed by sigma3A. Twenty-eight years on, the finding stands alone: PubMed's cited-in
link set for PMID:9792713 returns 5 citing papers (against 119 for PMID:14691137 and 70 for
PMID:21097499, same query), a PubMed search for "sigma3A adaptor" returns 2 papers total,
and UniProt's SUBUNIT line for Q92572 does not list IRS1. MARK_AS_OVER_ANNOTATED — an
unreplicated binding observation, not a demonstrated pathway role, and TAS rather than an
experimental code so there is no curator-read-the-full-text deference owed.

13. Affinage record

Trust gates clear (self_evaluation_pairwise: win, faith_pct: 100.0, and
.affinage.log reports "trust gates clear"). The record describes the right protein — it
opens on CLAPS3/AP3S1 as a clathrin-adaptor small chain and an AP-3 subunit, and its
UniProt accession field is Q92572 — so there is no AGT/AGXT-style symbol collision here.
The citation list is four numeric PMIDs, no bioRxiv ids. Graded MEDIUM / VERIFIED: the
four claims it makes are each traceable to their PMID, and I re-read all four.

Two cautions on its content, both checked:

• Its mechanism_profile proposes GO:0060090 molecular adaptor activity. Not imported.
  Re-grounded from the narrative and the PMIDs to GO:0140312 (§2.3).
• Two of its four findings are Low confidence and are association studies with no
  mechanism: an siRNA/co-IP virology screen PMID:38560106 and an ovarian-cancer
  oncodrive analysis PMID:38609993. Neither becomes an annotation. Neither does the
  pan-cancer expression paper PMID:35874816 or the cervical-carcinoma expression paper
  PMID:17125464 that I found separately.

What affinage missed — again, the whole molecular-function story. PMID:14691137,
PMID:21097499 and PMID:16162817 are the three papers that establish what sigma3A does, and
none appear in its citation list. Nor does the human AP-3 structure that contains this
protein (PMID:39705307, PDB 9C58/9C59/9C5B/9C5C), nor the delta-ear/sigma3 regulatory
mechanism (PMID:15469849), nor the subunit-assembly genetics (PMID:11807095), nor the
compartment paper (PMID:15051738). This is the documented failure mode exactly: the
decisive papers are titled for the complex, for the cargo, or for a virus, never for
AP3S1. Found instead by PubMed searches on "dileucine AND (AP-3 OR sigma3)", "sigma3A
adaptor", "AP-3 adaptor complex AND (cryo-EM OR crystal structure)" and by resolving the
PDB entries in the UniProt DR PDB; lines through the RCSB API, which named the PNAS
paper. Europe PMC's REST search returned HTTP 503 for the whole session, so all literature
search here went through NCBI E-utilities and RCSB instead.

14. Annotation dispositions

• AP-3 complex membership (GO:0030123 x3, GO:0030119, GO:0030117): ACCEPT. The GO term
  definition names sigma3A and sigma3B.
• Golgi apparatus, cytoplasmic vesicle membrane, transport vesicle, early endosome:
  ACCEPT for the first two (curated UniProt locations, vocabulary-mapped), KEEP_AS_NON_CORE
  for the last two.
• vesicle-mediated transport (IBA, IEA, NAS), protein transport, intracellular protein
  transport, clathrin-coated vesicle cargo loading: ACCEPT.
• Golgi to vacuole transport: MODIFY -> GO:0008333 (§6).
• synaptic vesicle coating: MODIFY -> GO:0016182 (§11).
• protein binding (IPI, tyrosinase): MODIFY -> GO:0140312 (§2.2).
• AP-3 adaptor complex binding: REMOVE (§7).
• insulin receptor signaling pathway: MARK_AS_OVER_ANNOTATED (§12).
• anterograde axonal transport x2, anterograde synaptic vesicle transport x2, axon
  cytoplasm, presynapse, synaptic vesicle recycling, membrane, melanosome assembly,
  platelet dense granule organization: KEEP_AS_NON_CORE (§10, §11).
• NEW: one row, GO:0005198 structural molecule activity, IDA, PMID:39705307 (§4). The
  more interesting subunit-level MF — dileucine sorting-signal binding — gets no NEW row
  because no GO term exists for it; that is the proposed_new_terms entry and the ONTOLOGY
  knowledge gap.

15. Final counts (all scripted, none asserted from memory)

• 30 GOA rows, 30 seeded review rows, reconciled one-to-one on
  (GO id, evidence code, reference, normalized WITH/FROM) by .scratch/reconcile.py.
• Actions, counted by .scratch/reconcile.py: ACCEPT 13, KEEP_AS_NON_CORE 12, MODIFY 3,
  MARK_AS_OVER_ANNOTATED 1, REMOVE 1, NEW 1 (31 rows in total).
• 18 rows carry WITH/FROM; all 18 have supporting_entities, and the 17 that the brief
  requires it for (1 IBA + 16 IEA/ISS) have a propagation_review; the IPI row carries one
  too, because the schema puts residue_claims inside that object. All 34 source_entities
  are generated from the seeded lists by .scratch/build_review.py rather than typed, so
  they cannot drift, and .scratch/reconcile.py asserts the two lists are equal per row.
• 32 references, every one with a reference_review; 168 supporting_text quotes, all
  verbatim per checkquotes.py.
• 3 residue claims, all RETAINED, all resolving against the live sequences.