AP3M1 (human, Q9Y2T2) — curation notes

Reviewer journal for the PAINT/affinage wave-B review. Provenance is inline as
[PMID:NNNN "verbatim quote"]. Everything asserted here is either quoted, computed in
AP3M1-bioinformatics/, or fetched from a named API during this session.

1. Identity check

AP3M1-uniprot.txt is ID AP3M1_HUMAN Reviewed; 418 AA. with
AC Q9Y2T2; Q5JQ12; Q9H5L2; — the expected accession, not a merged redirect. The record
names the protein "AP-3 complex subunit mu-1" with alternative names "AP-3 adaptor complex
mu3A subunit", "Mu-adaptin 3A", "Mu3A-adaptin". Single MHD domain, residues 176–417
(FT DOMAIN 176..417 / /note="MHD"), 418 aa, no catalytic features, no
active-site or binding features annotated.

PANTHER family from the UniProt DR line: PTHR10529 "AP COMPLEX SUBUNIT MU". This is a
single family spanning every AP medium subunit — the local
interpro/panther/PTHR10529/PTHR10529-entries.csv contains AP1M1, AP1M2, AP2M1, AP3M1,
AP3M2 and AP4M1 as human members — which is the structural reason several IBA rows on this
gene carry AP-1/AP-2-flavoured terms (§5).

InterPro matches (fetched from the InterPro API for Q9Y2T2, count 7): IPR001392
(Clathrin adaptor, mu subunit), IPR011012 (Longin-like domain superfamily), IPR018240
(Clathrin adaptor, mu subunit, conserved site), IPR022775 (AP complex, mu/sigma subunit),
IPR028565 (Mu homology domain), IPR036168 (AP-2 complex subunit mu, C-terminal
superfamily), IPR050431 (Adaptor complexes medium subunit). Exactly the seven listed on the
UniProt DR lines — the two lists agree, which matters for the ARBA analysis in §7.

2. Which AP-3 is this? AP-3A, not AP-3B

AP-3 is a heterotetramer. UniProt:
AP3B1 or AP3B2), a medium adaptin (mu-type subunit AP3M1 or AP3M2) and — i.e. two of the
four subunits come in ubiquitous and neuronal isoforms. Simpson et al. named them when they
characterised the complex: PMID:9151686
(p47A = mu3A = AP3M1; p47B = mu3B = AP3M2).

The ComplexPortal web service (queried this session) makes the pairing explicit. AP3M1 is a
subunit of exactly the two ubiquitous variants:

UniProt cross-references Q9Y2T2 to CPX-5051 and CPX-5052 only
(DR ComplexPortal; CPX-5051; Ubiquitous AP-3 Adaptor complex, sigma3a variant.). The
delta subunit AP3D1 is shared by all four, which is the pivot for §6: every mouse phenotype
that GOA transfers onto AP3M1 comes from the mocha (Ap3d1) allele, and a delta-null
removes the neuronal complex as well as the ubiquitous one.

GO has no term for either variant: GO:0030123 "AP-3 adaptor complex" has no children
(QuickGO /children returned none). Recorded as a proposed new term.

3. What the medium subunit does: YxxΦ cargo recognition

This is the part of AP3M1's biology that GOA does not capture at all, and it is the
best-evidenced thing about the protein.

The general principle was established for the family: PMID:9151686 and, in the same paper,
PMID:9151686.

Direct evidence on mu3A itself, four independent lines:

1. TGN38 YQRL. PMID:9118953 p47A is AP3M1.
2. Combinatorial specificity. PMID:9748267 and PMID:9748267
3. Selectivity for a lysosomal membrane protein. PMID:9794796 lgp120 is LAMP1.
   The same paper reports the AP-4 medium subunit behaving differently
   PMID:9794796
4. Structure. The 2024 human AP-3 cryo-EM series (PDB 9C58–9C5C, all cross-referenced
   from Q9Y2T2 with chain M = 1–418) resolves a LAMP1 YQTI peptide in the mu3 pocket:
   PMID:39705307

Note how well 3 and 4 agree: Stephens & Banting's lgp120 tail and Begley's LAMP1 peptide are
the same YQTI motif on the same protein, twenty-six years apart.

A fifth, independent line comes from virology: PMID:29028839 with the motif identified —
PMID:29028839

GOA carries none of this. The only molecular function on AP3M1 in the GOA file is
GO:0035615 (IBA, seeded from AP-1/AP-2 donors) plus GO:0031267 and 23 bare
GO:0005515 rows. The obvious comparator is the paralogue: human AP2M1 (Q96CW1) carries
GO:0005048 "signal sequence receptor activity" with IDA evidence from PMID:8918456
(QuickGO, queried this session), which is exactly the mu-chain YxxΦ activity —
PMID:8918456, measured
PMID:8918456. That paper is now
cached so the precedent is checkable from the repository and not only from a live query. So GO already
has a term and a curation precedent for this on the sister subunit, and AP3M1 simply lacks
the row. That is the main NEW recommendation of this review.

One bookkeeping note on that term. GO's current primary label for GO:0005048, per both OLS
and QuickGO queried this session, is "signal sequence receptor activity", with "signal
sequence binding" listed as an exact synonym. The repository's committed ontology cache
(cache/ontologies/go.tsv, fetched 2026-03-21) still carries "signal sequence binding", and
that cache is deliberately the validation authority here - its README states that cache
updates belong in separate commits from label fixes. The review therefore writes the cached
label, and this paragraph records the drift so it is visible rather than silent.

Bioinformatics: is the pocket actually there?

Rather than assert conservation, I measured it — AP3M1-bioinformatics/cargo_pocket.py,
results in AP3M1-bioinformatics/RESULTS.md. The script downloads the structures and
sequences live, asserts that PDB 9C5B chain M matches Q9Y2T2 at every one of its 418
author-numbered positions and that 1BXX chain A matches Q96CW1 at all 256 of its modelled
positions (so both are in UniProt numbering), then computes heavy-atom contacts at 4.5 Å.

• 9C5B chain M (AP3M1) vs the LAMP1 SHAGYQTI peptide: 10 contacts — E178, Y180, F181,
  V389, L392, F402, K403, G404, V405, K406.
• 1BXX chain A (AP2M1) vs the TGN38 DYQRLN peptide: 14 contacts — F174, L175, D176,
  K203, V401, R402, Y403, L404, V418, I419, K420, W421, V422, R423.
• 9 of AP3M1's 10 contacts sit on alignment columns that are also AP2M1 pocket positions.
  The site is the same site.
• But AP3M1 is identical to AP2M1 at only 5 of the 14 pocket columns (D182, V389, L392,
  K403, V405), while aligning to a residue at all 14. The substitutions include
  AP2M1 W421 → AP3M1 G404 and AP2M1 R402 → AP3M1 N390. Structurally conserved, chemically
  divergent — which is precisely what Ohno's combinatorial screen found functionally.
• Mouse Ap3m1 (Q9JKC8) is identical to human AP3M1 at 14 of 14 of these positions, so
  the ISS/IEA transfers GOA makes from mouse are not crossing a diverged cargo site.

These are encoded as machine-checkable residue_claims on the GO:0035615 row.

4. Is AP-3 a clathrin adaptor? No — and three GOA rows assume it is

This is the single biggest curation issue on the gene. The GOA file asserts clathrin three
times: GO:0030131 "clathrin adaptor complex" (IEA, InterPro2GO), GO:0035615
"clathrin-cargo adaptor activity" (IBA), and GO:0035654 "clathrin-coated vesicle cargo
loading, AP-3-mediated" (NAS, ComplexPortal).

The history: Simpson's group described AP-3 as a non-clathrin coat
PMID:9151686, and UniProt
still records that view (not clathrin-associated. The complex is associated with the Golgi). Dell'Angelica then reported the opposite PMID:9545220
— note that this is a beta3 activity, not a mu3A one. The 2024 structural review of the
field states the dispute as unresolved: PMID:39705307

The 2026 cryo-ET reconstitution settles the coat question: PMID:42139345 and the same paper places the complex
PMID:42139345 … on early/tubular
endosomes.

Ontology check (QuickGO, this session): GO:0030131 "clathrin adaptor complex" (def: "A
membrane coat adaptor complex that links clathrin to a membrane") has exactly two
children, GO:0030121 (AP-1) and GO:0030122 (AP-2). GO:0030123 (AP-3) is not among
its descendants — its ancestors are GO:0030119 AP-type membrane coat adaptor complex,
GO:0030117, GO:0048475, GO:0098796, GO:0005737. GO itself therefore places AP-3
outside the clathrin adaptor complexes, and the term definition says so:
GO:0030123 def "…AP-3 does not appear to associate with clathrin in all organisms."

So the InterPro2GO row GO:0030131 asserts, of AP3M1, membership in a class the ontology
explicitly does not put AP-3 in. The mapping fires because IPR001392 is "Clathrin adaptor,
mu subunit" — a family signature covering all AP medium subunits, AP-3 and AP-4 included.
That is a demonstrably wrong domain→complex mapping, which is the one licensed use of
REMOVE on an IEA row. (The same InterPro pair also yields GO:0006886 and GO:0016192,
which are generically true and stay.)

GO:0035615's definition is "Bringing together a cargo protein with clathrin, responsible
for the formation of endocytic vesicles" — wrong on both counts for AP-3 (no clathrin
lattice; not endocytic). Its parent GO:0140312 "cargo adaptor activity" is the right level
and is the proposed replacement.

GO:0035654's own definition already hedges ("in some organisms, links clathrin"). Its
grandparent GO:0035459 "vesicle cargo loading" carries the same assertion without the
clathrin commitment, so that is the proposed replacement, with a request for a
clathrin-neutral AP-3-specific child.

5. The five IBAs — nodes, seeds, and where the AP-1/AP-2 leakage is

just fetch-panther-paint PTHR10529 produced
interpro/panther/PTHR10529/PTHR10529-paint.tsv (10 nodes, 18 node-level annotations).
Every donor below was resolved against UniProt this session (xref: lookups) or against
PTHR10529-entries.csv.

Two nodes matter. PTN002237676 is the AP-3 mu clade node: both its seeds are genuine
AP-3 mu subunits, and AP3M1 is the human AP-3 mu subunit. That IBA is exactly right.
PTN000055849 is a deep node ancestral to the whole medium-subunit family (its
annotations reach AP3M1 and AP4M1), and three of the five IBAs hang off it.

GO:0035615 is the leakage case the family predicts. Every seed is an AP-1 or AP-2 medium
subunit — the two complexes that really are clathrin adaptors — and the term is then
inherited by AP-3 and AP-4 mu subunits, which are not. The current PAINT slice
(IBD dated 20260828) also lists UniProtKB:Q9Y6Q5 = human AP1M2 as a seed; the GOA
WITH/FROM predates that and lists three. Either way the seed set is 100% AP-1/AP-2. Human
AP4M1 (O00189) carries the same GO:0035615 IBA, and AP4M1's own experimental record
(QuickGO, this session) contains no clathrin activity at all.

Note also what the PAINT curators did fence off: PTN000242370 (the AP-2 mu clade) carries
an IRD (negated) for GO:0005802, and PTN002575694 (the stonin clade) carries an IRD
for GO:0006896. So the node placements have been actively curated; the objection to
GO:0035615 is that the clathrin-specific child was placed at a node whose descendants
include two non-clathrin coats, not that nobody was paying attention.

6. The mouse Ap3m1 record, and what GOA transfers from it

Five GOA rows on AP3M1 are IEA via GO_REF:0000107 (Ensembl Compara) with
UniProtKB:Q9JKC8|ensembl:ENSMUSP00000117346, and two more are ISS via GO_REF:0000024
with UniProtKB:Q9JKC8. Q9JKC8 is mouse Ap3m1 (UniProt lookup). Querying QuickGO for
Q9JKC8 (31 annotations) gives the experimental originals:

PMID:21998198 (Larimore et al., MBoC 2011) is the source of both "anterograde" rows. Its
AP-3 genetics are entirely the mocha delta allele: PMID:21998198, and the phenotype is
PMID:21998198. Searching the cached full text, the strings Ap3m1, mu3A and AP3M do not
occur; μ3 occurs twice, both in the introductory sentence naming the heterotetramer's
subunits. So mu3A is never perturbed or assayed — the evidence is complex-level, and since
delta is shared by the ubiquitous and the neuronal complex (§2), the phenotype cannot even
be assigned to AP-3A. Kept, but as non-core, with SOURCE_WEAK_OR_INFERRED.

That reading is reinforced by the genetics that do separate the two complexes:
PMID:15537701,
PMID:15537701 and PMID:15537701 — with the ubiquitous complex going the other
way: PMID:15537701. The synaptic-vesicle cargo job belongs
mostly to the neuronal complex, which does not contain AP3M1.

PMID:27568566 (Steinmetz et al., Cell Reports 2016) is the one paper that studies mu3A
itself in neurons, in mouse: PMID:27568566 and
PMID:27568566 synaptic scaling. Two caveats worth recording:
the sufficiency experiments are overexpression, and the paper's own conclusion is that the
effect is AP-3-independent ("excess μ3A acts independently of the AP-3A complex"). Kept as
non-core; SynGO curators read the full text and I am not second-guessing the IDA.

PMID:19010779 (Salazar et al., JBC 2009) cross-linked and affinity-purified AP-3 and
identified associated proteins by MS. Finding mu3A in a purified AP-3 preparation is
complex membership, which AP3M1 already asserts through GO:0030123 part_of; re-stating
it as the molecular function "AP-3 adaptor complex binding" types an obligate subunit as an
external ligand of its own complex. Marked as over-annotated rather than removed.

7. The other electronic rows

GO:0005737 cytoplasm comes from ARBA:ARBA00026971. Fetching
https://rest.uniprot.org/arba/ARBA00026971 returns a rule with 2388 condition sets and
a single output annotation, GO:0005737. Only two of those condition sets mention any
signature AP3M1 carries, and neither can fire for this protein: one requires IPR011012
AND taxon Saccharomyces; the other requires IPR022775 AND IPR016635 AND
IPR027156, and AP3M1's complete InterPro match list (§1, seven entries, cross-checked
against the InterPro API) contains neither IPR016635 nor IPR027156. So the row cannot be
reproduced from the rule's published conditions. This repo has already documented the same
pattern for ARBA00027853 (ARBA00027853-review.yaml). The annotation itself is harmless and
true — AP-3 is a cytosolic coat — so it is kept as non-core, with the reproducibility problem
recorded rather than used as grounds for removal (the API may not expose everything the
production pipeline uses).

GO:1904115 axon cytoplasm is GO_REF:0000108 — a logical inter-ontology inference whose
supporting_entities is the GO term GO:0008089. It is only as good as that row, which is
the mocha-derived one. Kept as non-core, flagged as inherited.

8. Reference-projection checks (QuickGO, paginated by entity)

PMID:17897319 is a 242-entity placental-lysosome membrane proteome
PMID:17897319; a cytosolic coat co-purifying with that fraction is
expected and non-core. PMID:22511774 and PMID:19116314 project over 7 and 4 entities
respectively — small, complex-scoped, and consistent with the papers' claims
(PMID:22511774). Neither is a mass projection.

PMID:10024875 deserves a note: the two TAS rows exist on AP3M1 and on no other protein in
GOA, even though the paper is about beta3A mutations in Hermansky-Pudlak syndrome. The
reason the curator chose mu3A is in the abstract — PMID:10024875 — so the process row
(GO:0006622) is well placed. The location row (GO:0005764 lysosome) is not: the paper
shows AP-3 sorts cargo to lysosomes PMID:10024875, which is a process, not residence in the organelle.

9. The bare protein binding rows

Twenty-three GO:0005515 IPI rows. Twenty-two come from six high-throughput interactome
papers — HuRI/Y2H (PMID:25416956, PMID:32296183), BioPlex AP-MS (PMID:28514442,
PMID:33961781), a neurodegeneration Y2H map (PMID:32814053) and the SLC interactome
(PMID:40355756) — with partners AGTRAP, ARNT2, CRMP1, DVL3, FRMD6, HLA-DMB, HTT, IL36RN,
RSPH14, SIPA1L2, SLC12A4, SPAG16, TRIM9, TRIM23. None has functional follow-up naming
AP3M1, and none of them is an AP-3 subunit or a known AP-3 cargo whose sorting was traced.
The UniProt NbExp values on these pairs (2–13) are replicate counts inside those same
datasets, not independent studies. All marked over-annotated; a bare protein binding row
adds nothing that the informative MF row proposed in §3 does not say better.

The twenty-third (UniProtKB:Q04671 = OCA2, PMID:19116314, assigned by UniProt rather than
IntAct) is a genuine low-throughput result, but the abstract is all the cache has and the
motifs assayed are dileucine, not tyrosine: PMID:19116314. Acidic dileucine signals are read by the
sigma3/delta hemicomplex, not by the mu3A tyrosine pocket
(PMID:39705307 — the dileucine site is on sigma). With
the full text unavailable I cannot tell which construct was used, so this stays as an
over-annotated bare binding row rather than being converted to a cargo-recognition MF.

The two GO:0031267 rows (RAB38 P57729, RAB32 Q13637; PMID:22511774) are the melanocyte
co-immunoprecipitation above. Real, but a complex-level co-IP in a specialised cell type;
kept as non-core.

10. Affinage record

AP3M1-deep-research-affinage.md has self_evaluation_pairwise: loss, faith_pct: 50.0,
n_discoveries: 1, citation_count: 1, and .affinage.log reports the trust gate tripped.
Its entire content is one 2017 proteomics correlation — PMID:29032074 — i.e. that AP3B1 depletion
lowers AP3M1 abundance. That is a real and correctly-cited observation about complex
stoichiometry, and it is the only thing in the record; its own summary concedes "no further
mechanistic detail for AP3M1 has been characterized in the available corpus", which is
plainly false for this gene.

What affinage missed (every one of these was found by independent Europe PMC search):
the entire YxxΦ literature (PMID:9118953, PMID:9748267, PMID:9794796, PMID:11071885), both
AP-3 cryo-EM papers (PMID:39705307, PMID:42139345) — the first of which is cross-referenced
straight from the UniProt record via PDB 9C58–9C5C — the clathrin-association dispute
(PMID:9545220 vs PMID:9151686), the HPS2 cargo phenotype (PMID:10024875), the mu3A neuronal
paper (PMID:27568566), the beta3A/beta3B genetic dissection (PMID:15537701), and the HRSV
matrix interaction (PMID:29028839) that UniProt itself cites as reference [6]. Marked
LOW_QUALITY / relevance: LOW in references.

11. Things that stayed open

• Is human AP3M1 associated with any disease? Searches on "AP3M1" AND (mutation OR variant OR patient) (143 hits) returned no Mendelian AP3M1 disorder; the human AP-3 diseases are
  HPS2 (AP3B1) and HPS10 (AP3D1) PMID:31898847, and the 2020 HPS mutation update states that all ten subtypes have an assigned
  gene PMID:31898847. The string AP3M1 does not occur anywhere in
  that review's cached full text. AP3M1 shares a bidirectional CpG-island promoter with
  adenosine kinase PMID:14575525, which is a plausible reason large ADK-region deletions could
  take AP3M1 with them, but no human case has been reported.
• Whether mu3A and mu3B genuinely partition with beta3A and beta3B respectively, or "mix and
  match", was an open question when AP-3 was characterised and I found nothing settling it
  for human cells.
• Whether the residual clathrin association of AP-3 (beta3 hinge, PMID:9545220) has any
  in-cell function, given that carrier formation does not need a clathrin lattice
  (PMID:42139345).