AP3M2 (mu3B-adaptin) — curation notes

UniProt P53677 (AP3M2_HUMAN, "AP-3 complex subunit mu-2", 418 aa, chromosome 8,
HGNC:570). Verified the accession returns the intended protein rather than a merged
record: the fetched entry is ID AP3M2_HUMAN Reviewed; 418 AA.
with GN Name=AP3M2;. PANTHER family PTHR10529 "AP COMPLEX SUBUNIT MU"
(DR PANTHER; PTHR10529; AP COMPLEX SUBUNIT MU; 1.).

GOA snapshot: 26 rows, of which 4 IBA, 2 ISS, 11 IEA, 4 NAS, 1 TAS — and no
experimental rows at all. So this review is almost entirely a propagation audit.
A live QuickGO pull for P53677 at review time returned the same 26 rows as the
cached AP3M2-goa.tsv, so the snapshot is current.

One snapshot divergence worth flagging but not actionable: the UniProt record's own
DR GO block lists GO:0006897; P:endocytosis; IBA:GO_Central and the entry says
DR PAN-GO; P53677; 5 GO annotations based on evolutionary models., whereas GOA
carries only 4 IBAs and no GO:0006897. The PTHR10529 PAINT slice fetched for this
review carries no GO:0006897 IBD node either. No row was added for it: the review
follows GOA.

1. What the protein is

AP3M2 is one of two mammalian mu3 medium subunits of the heterotetrameric AP-3
adaptor. UniProt states the composition explicitly:

CC Adaptor protein complex 3 (AP-3) is a heterotetramer composed of two
CC large adaptins (delta-type subunit AP3D1 and beta-type subunit AP3B1 or
CC AP3B2), a medium adaptin (mu-type subunit AP3M1 or AP3M2) and a small
CC adaptin (sigma-type subunit APS1 or AP3S2). {ECO:0000250}.
— file:human/AP3M2/AP3M2-uniprot.txt (quoted line-by-line: a file: supporting_text must sit on one physical line)

The two mu3 paralogues define the two AP-3 isoforms. AP-3A is the ubiquitous complex
(beta3A/AP3B1 + mu3A/AP3M1); AP-3B is the neuronal complex (beta3B/AP3B2 +
mu3B/AP3M2), with the delta and sigma3 subunits shared. Nakatsu et al. state the
pairing and, importantly, that it is obligate:

• PMID:15492041
• PMID:15492041

ComplexPortal encodes exactly this: CPX-5053 "Neuronal AP-3 Adaptor complex, sigma3b
variant" and CPX-5055 "…sigma3a variant" both list AP3D1 + AP3B2 + AP3M2 +
AP3S2/AP3S1. Both ComplexPortal records are cross-referenced from the UniProt entry.
This matters for the review because every ComplexPortal NAS row on AP3M2 is a
complex-level statement projected onto the subunit.

The historical literature calls the protein p47B. Pevsner et al. cloned it from rat
and reported the expression pattern that gave rise to the "neuron-specific" framing
PMID:8076832.
Simpson et al. restated it when naming AP-3
PMID:9151686.

But human AP3M2 is not neuron-restricted

The "exclusively brain and spinal cord" statement is a 1994 rodent Northern blot. The
human gene behaves differently. Human Protein Atlas (fetched live, see
AP3M2-bioinformatics/RESULTS.md §5) reports RNA tissue specificity: Low tissue
specificity, RNA tissue distribution: Detected in all, with single-cell
enrichment in melanocytes (103.5 nCPM) and spermatogenic cell types rather than in
neurons; the single-nuclei brain data show low cell-type specificity. The UniProt
entry agrees (DR HPA; ENSG00000070718; Low tissue specificity.;
DR Bgee; ENSG00000070718; Expressed in endothelial cell and 186 other cell types or tissues.).
The tissue-level expression cluster is still "Brain & retina - Neuronal signaling", so
the gene is brain-weighted without being brain-restricted.

The pigment-cell enrichment is not an artefact of one atlas. Zebrafish ap3m2 is
required for lysosome-related-organelle biogenesis in iridophores
PMID:41950095
and ap3m2 was one of the candidates in the transparent zebrafish pinky mutant
PMID:23639161.
Human cell work outside neurons also exists: AP3M2 knockdown reduces IL-6 secretion in
iPSC-derived astrocytes PMID:30371777
and reduces viability of colorectal-cancer lines PMID:39488930.
I did not convert any of this into annotation actions — the zebrafish gene's
orthology to human AP3M2 rather than AP3M1 was not established here, and the human
knockdown phenotypes are not mechanistically resolved. It is recorded as a knowledge
gap instead.

2. What AP-3B does

2.1 It buds synaptic vesicles from endosomes — and that is the mu3B-specific job

The decisive paper is Blumstein et al. 2001, which affinage did not return (see §6):

• PMID:11588176
• PMID:11588176

It builds on the original AP-3 coat reconstitution
[PMID:9590176 "Depletion of AP3 from brain cytosol inhibits small vesicle formation from PC12 endosomes in vitro."; PMID:9590176 "We conclude that AP3 coating is involved in at least one pathway of small vesicle formation from endosomes."]
and is confirmed in vivo by the mu3B knockout
PMID:15492041.

Nakatsu et al. also draw the distinction that the GO terms in GOA get wrong:

PMID:15492041

and place AP-3B on the second of the two
PMID:15492041.

GO has a term for exactly this: GO:0016182 synaptic vesicle budding from endosome,
"Budding of synaptic vesicles during the formation of constitutive recycling vesicles
from early endosomes" (QuickGO). Mouse Ap3d1 already carries it with IMP and IDA from
PMID:11588176. Human AP3M2 carries neither it nor anything under it; instead it carries
two terms on the endocytic branch — GO:0016183 synaptic vesicle coating (whose
definition is "The formation of clathrin coated pits in the presynaptic membrane
endocytic zone…") and GO:0048488 synaptic vesicle endocytosis ("…the synaptic vesicle
membrane constituents are retrieved from the presynaptic membrane on the axon terminal
after neurotransmitter secretion by exocytosis"). Checked with QuickGO: GO:0016183
is_a GO:0048488 and both sit under GO:0006897 endocytosis, whereas GO:0016182 does
not. Both rows are therefore MODIFY → GO:0016182.

2.2 The in vivo phenotype is a selective GABAergic defect

The mu3B knockout is the only gene-specific in vivo evidence that exists:

• seizures — PMID:15492041
• selective transmitter defect — PMID:15492041
• the cargo — PMID:15492041
• ultrastructure at both synapse types — PMID:15492041

This is why GO:0098982 GABA-ergic synapse is graded as core while GO:0098978
glutamatergic synapse is kept as non-core: both compartments show a vesicle-density
phenotype, but only the inhibitory one shows a functional transmitter defect.

Human genetics has not (yet) matched the mouse: a 190-patient screen found
PMID:17293072.
The obligate partner AP3B2, by contrast, causes DEE48
PMID:41948612.
Mouse Ap3m2 is also the positional candidate at an alcohol-preference/withdrawal QTL
PMID:24923803 —
recorded as context, not annotated.

2.3 Where AP-3 acts: endosomes, not the TGN

This was a genuine controversy and it was settled by immuno-EM:

• PMID:15051738
• PMID:15051738

GO's own definition of the complex agrees: GO:0030123 is "A heterotetrameric AP-type
membrane coat adaptor complex that consists of beta3, delta, mu3 and sigma3 subunits
and is found associated with endosomal membranes" (QuickGO). AP-3 is also seen on
early/recycling endosome tubules in pigment cells
PMID:23247405.

The earlier literature that put AP-3 "at the Golgi" was itself hedged
PMID:9151686,
and UniProt still records SUBCELLULAR LOCATION: Golgi apparatus. Cytoplasmic vesicle membrane. So Golgi is kept (non-core) while the IBA-derived is_active_in trans-Golgi network is modified to early endosome — see §4.

2.4 The molecular function: YxxPhi cargo recognition by the mu subunit

The mu subunits of AP complexes are the tyrosine-signal-recognition arm
PMID:9151686.
mu3B itself was directly assayed — this is the one molecular-function experiment on
this protein anywhere in the literature I found:

PMID:9748267
PMID:9748267

and the structural mechanism is now solved for the human AP-3 holocomplex
PMID:39705307,
with the cargo being the LAMP1 tail
PMID:39705307.
Note that mu3 is not the only cargo arm: the sigma3/delta hemicomplex carries the
dileucine site PMID:39705307,
and Nakatsu's candidate VGAT signal is a dileucine, not a YxxPhi
PMID:15492041.
So mu3B contributes one of two cargo-recognition sites in the complex, which is why
contributes_to is the right qualifier in core_functions.

GO has no term for this. Searching QuickGO for "sorting signal binding",
"tyrosine-based sorting", "dileucine" and "YXXphi" returns exactly one relevant
molecular-function term, GO:0089710 endocytic targeting sequence binding, whose
definition is explicitly committed to internalisation: "Binding to a endocytic signal
sequence, a specific peptide sequence, of 4-6 amino acids with an essential tyrosine
(Y), found on cytoplasmic tails of some cell surface membrane proteins, which directs
internalization by clathrin-coated pits." Its only ancestors are GO:0005515 /
GO:0005488 / GO:0003674, so there is no non-endocytic sibling to fall back on. That the
motif's recognition by a mu subunit is not an endocytic-only activity is shown directly
PMID:11139587.
Recorded under proposed_new_terms.

3. Bioinformatics (see AP3M2-bioinformatics/RESULTS.md)

Everything below is recomputed live by AP3M2-bioinformatics/analyze.py; nothing is
hardcoded, and wrong accessions raise rather than pass silently (the first run caught
my own use of P53676 — rat Ap3m1, not human).

Paralogy. AP3M2's closest human paralogue is AP3M1 at 84.2% identity; identity
to the clathrin-adaptor medium subunits is AP1M1 29.7%, AP1M2 31.6%, AP2M1 29.1%, and
to AP4M1 24.5%. The AP-3 pair is a clean, well-separated subfamily. This is the
quantitative form of the objection in §4 to a family-root IBD seeded only by AP-1/AP-2
donors.

Is the cargo site intact in mu3B? Defined empirically from PDB 9C5B, the human
AP-3 holocomplex on a nanodisc with the LAMP1 cytoplasmic tail bound (SIFTS maps chain
M to Q9Y2T2/AP3M1 1-418 and chain Y to P11279/LAMP1 406-417; modelled peptide
SHAGYQTI, carrying the GYQTI YxxPhi motif). Nine mu3A residues have a heavy atom
within 4.0 A of the cargo: Y180, F181, V389, L392, F402, K403, G404, V405, K406.
AP3M2 is identical at 8 of the 9, the ninth being a conservative V405I. No
deletions. There is therefore no residue-level evidence that the neuronal paralogue has
lost cargo recognition — encoded as residue_claims on the molecular-function row.

A second, independent AP-3 cargo complex. The PANTHER family metadata names
PDB 4IKN as PTHR10529's representative structure: the rat mu3A C-terminal domain
bound to the TGN38 tail (DYQRL) — a different species and a different cargo from 9C5B.
Ten rat mu3A residues contact the peptide, and projected onto human AP3M1 nine of them
are exactly the 9C5B set; the tenth is D182, which falls just outside the 4.0 Å cutoff
in 9C5B. Human AP3M2 carries the same residue as rat mu3A at 9 of the 10. So the site is
not an artefact of one structure, one cargo or one species, and mu3B matches mu3A across
it. (D182 is one of only two positions at which every human mu paralogue matches mu2 in
the outgroup comparison below — the other is mu2 V401, which maps to V389 and is itself a
contact in both AP-3 structures. analyze.py computes that count rather than my asserting
it; every other mu2 contact position has at least one paralogue that diverges.)

Outgroup and alignment check. The classical AP-2 pocket from PDB 1BXX (rat mu2 +
TGN38 DYQRLN) gives 13 contact residues. Both AP-3 mu subunits have diverged sharply
from it — AP3M2 matches mu2 at 4/13 and AP3M1 at 5/13, versus AP1M1 10/13, AP1M2 9/13
and AP4M1 11/13 — most strikingly mu2 W421 → G404 in both mu3 proteins. Yet all 13
positions still align with no deletion, and projecting them onto AP3M1 lands on 9 of
the 9 positions 9C5B observes contacting LAMP1. The two routes agree, so the
cross-family alignment is not drifting, and the reading is that AP-3 recognises
tyrosine cargo through the structurally equivalent site with a diverged residue
complement — which is precisely what Ohno et al. reported functionally
PMID:9748267.

mu3 linker amphipathic helix. Begley et al. report a membrane-inserting amphipathic
helix in the mu3 linker PMID:39705307.
Scanning the 45 residues preceding each protein's own UniProt MHD boundary with an
Eisenberg hydrophobic-moment window, AP3M2 139-156 (ILRTVVNTITGSTNVGDQ) scores
= 0.487 against AP3M1 0.450, AP1M1 0.433, AP1M2 0.373, AP2M1 0.227 and AP4M1
0.161 — the two mu3 proteins hold the top two moments. Mean hydrophobicity, by contrast,
is positive for AP3M2 (+0.10), AP3M1 (+0.12) and AP4M1 (+0.44), so ⟨H⟩ alone does not
separate mu3 from the rest and it is the moment that does; analyze.py computes that
list rather than my eyeballing the table. Consistent with mu3B retaining the mu3
membrane-insertion feature. This is a prediction, not an observation, and is reported as
such.

Isoform 2. P53677-2 combines VAR_SEQ 268-273 (NLVAIP -> KCCLGM) with VAR_SEQ
274-418 (Missing), i.e. it truncates at residue 273 of 418 and so removes the
C-terminal half of the MHD (176-417). Computed against the cargo-contacting positions
rather than assumed: the truncation deletes seven of the nine (389, 392, 402, 403,
404, 405, 406) and retains two (Y180, F181), which lie N-terminal to it — so isoform 2
keeps the start of the site and loses the part that reads the Ø position. Worth noting because the four IntAct interactions in
the UniProt record (FGFR3, HRAS, MEOX2, SPRED1) are all on the -2 isoform with
NbExp=3, which is replicate counts within one dataset, not three studies. No GO row
depends on them and none was added.

4. Propagation audit

4.1 The PANTHER family and its nodes

just fetch-panther-paint PTHR10529 →
interpro/panther/PTHR10529/PTHR10529-paint.tsv, 10 nodes / 18 node-level
annotations. PTHR10529 is broad: PTHR10529-entries.csv lists AP-1 mu, AP-2 mu, AP-3
mu, AP-4 mu and stonins (STON1 Q9Y6Q2, STON2 Q8WXE9, Drosophila stoned-B) among its
60 reviewed members. The node structure that matters:

AP3M2 receives IBAs from exactly two of these — PTN002237676 (the AP-3 clade) and
PTN000055849 (the family root) — and receives no AP-1/AP-2/AP-4 complex term, so the
complex-membership propagation is clean. The leakage is entirely at the root node.

Donors resolved through xref:<db>-<id> UniProt searches (fetch size 5; the only
multi-hit was AT4G24550 → Q9SB50 Swiss-Prot plus two TrEMBL isoforms):

4.2 GO:0035615 clathrin-cargo adaptor activity — the clearest defect

The PTN000055849 IBD for GO:0035615 is seeded by FB:FBgn0024833 (fly AP-1mu),
FB:FBgn0263351 (fly AP-2mu), UniProtKB:Q9Y6Q5 (human AP1M2) and
dictyBase:DDB_G0289247 (Dicty AP-1 mu). All four donors are AP-1 or AP-2 medium
subunits — the two clathrin adaptors — and there is no AP-3 or AP-4 donor at all.
The term's definition is doubly committed: "Bringing together a cargo protein with
clathrin, responsible for the formation of endocytic vesicles" (QuickGO). AP-3 is
neither: UniProt says Part of the AP-3 complex, an adaptor-related complex which is not clathrin-associated and Peden showed its budding profiles arise from endosomes, not
the plasma membrane. (AP-3 does contact clathrin — PMID:9545220 — but that is the beta3 subunit and it does not make AP-3 an endocytic coat.)
→ MODIFY to GO:0140312 cargo adaptor activity, the immediate parent (QuickGO
ancestors of GO:0035615 include GO:0140312, GO:0030674, GO:0060090), which keeps the
adaptor claim and drops the clathrin/endocytosis commitments.

This is not an AP3M2 quirk. Querying QuickGO for the same term on the other two human
non-clathrin medium subunits returns the identical row on both:

AP3M1  GO:0035615  IBA  GO_REF:0000033  FB:FBgn0024833|FB:FBgn0263351|PANTHER:PTN000055849|dictyBase:DDB_G0289247
AP4M1  GO:0035615  IBA  GO_REF:0000033  FB:FBgn0024833|FB:FBgn0263351|PANTHER:PTN000055849|dictyBase:DDB_G0289247

Same node, same four AP-1/AP-2 donors, same clathrin-committed molecular function on
three human genes whose complexes are not clathrin coats. Fixing it at the node would fix
all three. The residue analysis in
§3 is attached here so it is clear the modification is a scoping fix and not a
loss-of-function argument.

4.3 GO:0005802 trans-Golgi network and GO:0006896 Golgi to vacuole transport

Both also come from the root node. The TGN IBD is seeded only by AP-1 mu (AT1G60780,
plus yeast APM2) and AP-4 mu (AT4G24550, E2RED8, O00189) — no AP-3 donor — and the
PAINT curator has already placed an IRD on GO:0005802 at the AP-2 clade
(PTN000242370), showing the compartment is understood to be subfamily-specific within
this family. Mammalian AP-3's compartment is the tubular sorting endosome (§2.3).
→ MODIFY to GO:0005769 early endosome.

GO:0006896 does have a real AP-3 donor (yeast APM3), and in fungi AP-3 genuinely runs
the Golgi→vacuole ALP route. But that is the fungal route; the metazoan AP-3 route runs
endosome→lysosome, which is the finding of PMID:15051738. → MODIFY to GO:0008333
endosome to lysosome transport ("The directed movement of substances from endosomes
to lysosomes", QuickGO). Note this is the same node whose GO:0006896 assertion the
PAINT curators already blocked with an IRD one clade over (PTN002575694, stonins).

4.4 GO:0030123 AP-3 adaptor complex — sound

PTN002237676 is seeded by yeast APM3 and Dicty apm3, both bona fide AP-3 mu subunits,
and AP3M2 sits squarely inside the AP-3 mu clade (84% identical to AP3M1). AP3M2 is not
among its own donors, which is expected — the gene has no experimental GO row of any
kind. ACCEPT, NO_FAILURE_CORE.

4.5 The Ensembl-Compara IEA block (GO_REF:0000107) and the two ISS rows

Seven rows transfer from mouse Q8R2R9 (Ap3m2) and one from rat P53678 (Ap3m2).
Resolved both via UniProt REST (Q8R2R9 = Mus musculus Ap3m2, P53678 = Rattus
norvegicus Ap3m2, both Swiss-Prot, both PTHR10529). Pulling the donors' own GO records
from QuickGO shows what each transfer actually rests on:

Two of these do not survive inspection of the cited paper.

PMID:21998198 does not perturb Ap3m2. The cached full text (24,083 words) contains
zero occurrences of "Ap3m2", "mu3B", "μ3B" or "AP-3B"; "μ3" appears once, in the generic
sentence naming the complex's four subunits. Every mouse allele in the study is
BLOC-1 or the delta subunit: PMID:21998198.
Because mocha removes the shared delta subunit it ablates both AP-3A and AP-3B, so
the result cannot be assigned to the mu3B-containing complex. The biology is real and
AP-3 is genuinely required for cell-body-to-neurite cargo delivery
PMID:21998198,
so the four affected rows (GO:0008089 IEA + ISS, GO:0048490 IEA + ISS) are kept as
non-core rather than removed, with SOURCE_WEAK_OR_INFERRED recorded. I am not
asserting the mouse annotation is wrong — I did not see what the UniProt curator saw
beyond this text — only that the evidence behind the human transfer is complex-level.
GO:1904115 axon cytoplasm is a GOC logical inference from GO:0008089
(WITH/FROM = GO:0008089) and inherits the same status.

PMID:33376223 names AP-3 as an SV visitor, not a resident. The rat is_active_in GO:0008021 rests on an SV-fraction proteome whose authors write
PMID:33376223.
Their per-protein dataset (Dataset S1) is not in the cached text, so I cannot see
AP3M2's own rank or resident/visitor call. Kept as non-core with the caveat recorded;
Blumstein's independent in vitro result that neuronal AP-3 binds purified synaptic
vesicles is the reason it is kept at all.

GO:0035651 AP-3 adaptor complex binding is a role conflation at the term level:
AP3M2 is a constitutive subunit of AP-3, which GO expresses as part_of GO:0030123
(already present twice on this gene), not as an enzyme-style "binding" of its own
complex. The mouse IDA is abstract-only in our cache (full_text_available: false), so
I make no claim about what the mouse experiment showed;
MARK_AS_OVER_ANNOTATED applies to the human transfer only.

4.6 Historical GO:0030131 clathrin adaptor complex assessment (superseded)

The InterPro API gives IPR001392 "Clathrin adaptor, mu subunit" → GO:0006886,
GO:0016192 and GO:0030131. GO:0030131 is "A membrane coat adaptor complex that
links clathrin to a membrane". QuickGO's ancestor list for GO:0030123 is
GO:0005575, GO:0005622, GO:0005737, GO:0016020, GO:0030117, GO:0030119, GO:0030123, GO:0032991, GO:0048475, GO:0098796, GO:0110165 — GO:0030131 is not in it. So in
GO's own structure the AP-3 complex is not a clathrin adaptor complex, and this is not
a case of an IEA being harmlessly broader: it is a different branch. The mapping is
calibrated on mu1/mu2 and over-reaches to the mu3/mu4 members of the same InterPro
family. → MODIFY to GO:0030123. The other two InterPro2GO terms (GO:0006886,
GO:0016192) are correct for the whole family and are accepted.

4.7 ARBA00026971 (GO:0005737 cytoplasm) cannot be reproduced

https://rest.uniprot.org/arba/ARBA00026971 returns a rule with 2,388 condition
sets conferring a single annotation, GO:0005737. Filtering those sets for any of
AP3M2's seven InterPro signatures (IPR001392, IPR011012, IPR018240, IPR022775,
IPR028565, IPR036168, IPR050431, confirmed via the InterPro API for P53677) or for
PTHR10529 leaves exactly two:

• one pairing IPR011012 with taxon Saccharomyces — AP3M2 is human, so this fails;
• one pairing IPR022775 with IPR016635 and IPR027156 — IPR016635 is "Adaptor
  protein complex, sigma subunit" and IPR027156 is "AP-2 complex subunit sigma";
  AP3M2 is a mu subunit and has neither, so this fails too.

No condition set of the rule as the API serves it is satisfied by this protein. (The
record returns no meaningful version string, so an earlier revision may have had a set
that did fire; the finding is about the rule as it now stands.) The claim is
nonetheless true (AP-3 is a cytosolic coat that cycles on and off membranes), so the
original row was kept as non-core rather than removed. The recovery reassessment now
ACCEPTs the location as integral to coat recruitment; the historical rule-reproducibility
concern remains in propagation_review, with the version limitation explicit. It does not
establish that the original annotation was generated incorrectly. The earlier audit
compared this with the repo's ARBA00027853 finding.

5. Historical action summary and rationale index (superseded by the current YAML)

This table records the earlier audit, not the recovery decisions. In particular, cytoplasm
is now ACCEPT and clathrin adaptor complex is KEEP_AS_NON_CORE. See the current YAML
for all final actions. The earlier arguments are retained for traceability.

The TAS row (#12) deserves a word. PMID:8076832 is the 1994 cloning paper; it predates
the identification of AP-3 by three years and its own claim is homology-level
PMID:8076832.
The parent term it carries is correct — GO:0030119's definition explicitly covers AP-3,
"Any of several heterotetrameric complexes that link clathrin (or another coat-forming
molecule, as hypothesized for AP-3 and AP-4) to a membrane surface" — but the specific
complex identity has been known since 1997 PMID:9151686 and is what should be
asserted.

6. What affinage missed

The affinage record (gates_passed, self_evaluation_pairwise: win, faith_pct: 100.0, 6 citations) describes the right protein and gets the LRO/pigment-cell angle —
which turned out to be more useful than expected, given the HPA melanocyte enrichment.
But it returned none of the papers this review actually rests on. Missing:

• PMID:11588176 Blumstein 2001 — the only paper that isolates the neuronal AP-3
  function, and the single most important reference for AP3M2;
• PMID:15492041 Nakatsu 2004 — the mu3B knockout itself, already in GOA via the
  mouse ortholog, and the source of the GABAergic phenotype;
• PMID:9748267 Ohno 1998 — the only direct molecular-function assay of mu3B;
• PMID:9590176 Faundez 1998, PMID:15051738 Peden 2004, PMID:39705307
  Begley 2024, PMID:9151686, PMID:9118953, PMID:15537701.

The pattern is the one this campaign keeps hitting: the decisive papers are titled for
the complex ("AP-3", "AP-3B", "adaptor medium chains") or for the partner
(AP3B2/DEE48), not for the gene symbol, and a symbol-anchored search does not reach
them. Affinage did return the two 2024-2026 papers that are titled "AP3M2" — the CRC
knockdown and the zebrafish iridophore work — which is consistent with its gates
measuring precision rather than recall. Everything above was recovered by hand from
Europe PMC (which 503'd intermittently and needed a retry loop), from QuickGO donor
records, and from the PAINT slice.

7. Open questions

• Is human AP3M2 functionally neuronal at all? The rodent restriction (PMID:8076832) is
  not reproduced by human expression atlases, which put its highest single-cell signal
  in melanocytes and spermatogenic cells. Either the paralogue's tissue distribution
  changed in the primate lineage, or the protein has an unrecognised role in
  lysosome-related-organelle biogenesis in pigment cells, as the zebrafish knockout
  suggests.
• Does mu3B-containing AP-3B select different cargo from mu3A? Ohno's library screen
  says the YxxPhi preferences of mu3A and mu3B differ, but no mu3B-specific cargo has
  been identified; VGAT, the best candidate, is proposed to use a dileucine signal,
  which is read by sigma3/delta rather than by mu3.
• Is the AP-3B complex ever assembled without beta3B? Nakatsu's immunoblot says beta3B
  is destabilised in mu3B-null brain, implying strict interdependence, but the reverse
  experiment (mu3B stability in AP3B2-null neurons) was not found here — and it matters
  for interpreting DEE48.
• Which complex does the alcohol-preference QTL act through? PMID:24923803 maps
  regulatory variation upstream of Ap3m2 rather than coding change, so the phenotype may
  be a dosage effect on AP-3B assembly.

8. Validation record

Counts below are printed by AP3M2-bioinformatics/reconcile_goa.py, not asserted by
hand. That script is the GOA-to-YAML reconciliation the campaign brief requires: it
checks that every line of AP3M2-goa.tsv maps to exactly one existing_annotations
entry with the same term, evidence code, reference and normalised WITH/FROM set, that
the only extra entry is the NEW row, and that every propagation_review's
source_entities are exactly the row's supporting_entities, in order, each with a
comment.

GOA tsv rows            : 26
YAML reviewed rows      : 26
YAML NEW rows           : 1
rows with propagation_review: 21
rows with supporting_entities: 21
action counts           : {'KEEP_AS_NON_CORE': 11, 'ACCEPT': 7, 'MODIFY': 7,
                           'MARK_AS_OVER_ANNOTATED': 1, 'NEW': 1}
OK: GOA and review reconcile exactly

• just validate human AP3M2 → ✓ Valid (with 1 warnings). The single warning is that
  no annotation cites the affinage file in supported_by. That is deliberate: the
  campaign brief forbids quoting an affinage sentence as supporting_text for a
  mechanistic claim, and there is no non-mechanistic row here where the file would be
  the right source. The record is still adjudicated in references with
  relevance: LOW / correctness: LOW_QUALITY and an explanation. The campaign's own
  model review, genes/human/AGT, carries the identical warning with status: COMPLETE.
• checkquotes.py → checked 80 quotes: 0 failures, 0 skipped.
• uv run python -m ai_gene_review.validation.gene_residue_claims →
  352 pass, 0 fail, 0 unresolved over 20 gene reviews, AP3M2's five claims among them.
• Duplicate-key scan of the YAML → none. cache/go/terms.csv → no deletions relative to
  the merge base, no duplicate ids.

9. Changes made across the three bot review rounds

The third clathrin-committed term. The first draft modified GO:0035615 and
GO:0030131 because each commits AP-3 to clathrin, and then accepted GO:0035654
"clathrin-coated vesicle cargo loading, AP-3-mediated" without comment. That reads
as selective, and the objection does apply to the wording: the definition's clause is
"transported by a clathrin-coated vesicle". The distinction that makes ACCEPT right
anyway is about what each term picks out and what alternatives exist.

So the remedy for GO:0035654 belongs to the ontology, not to this gene: drop
"clathrin-coated" from the name and definition and re-parent it directly under GO:0035459
vesicle cargo loading, leaving its current parent GO:0035652 for the AP-1/AP-2 adaptors.
Both of those ids and labels were checked against QuickGO rather than written from memory
— GO:0035652 is "clathrin-coated vesicle cargo loading", not obsolete, with ancestors
GO:0006810, GO:0008150, GO:0035459, GO:0035652, GO:0051179, GO:0051234, and its
children are exactly GO:0035653 and GO:0035654, both is_a. That last lookup is the
one that establishes the relation the proposal turns on: GO:0035654 currently sits
beneath GO:0035652, so the proposal moves it up one level to sit alongside it. (Both proposed_new_terms entries are really ontology change requests, and
ProposedOntologyTerm has no field that says so; each justification now states it
explicitly, and the schema gap is worth an issue of its own.) That is now the second
entry in proposed_new_terms, and the ONTOLOGY knowledge gap names both defects as two
faces of the same legacy framing. The structural work states the point plainly
PMID:39705307.

Cross-review inconsistency with AP3B2 — not resolved here.
genes/human/AP3B2/AP3B2-ai-review.yaml ACCEPTs both GO:0016183 (same ComplexPortal
NAS, same PMID:15537701) and GO:0048488 (same Ensembl route), and puts GO:0016183 in
its core functions. This review modifies both to GO:0016182 for the obligate partner of
the same complex, so the repository now carries two opposite calls on the same assertion.

Worth recording that AP3B2's own summary for GO:0016183 reads "mediating the formation
of synaptic vesicle precursors from endosomal membranes" — which is the definition of
GO:0016182, not of GO:0016183 ("the formation of clathrin coated pits in the
presynaptic membrane endocytic zone"). The two reviews therefore agree on the biology and
differ only on which term expresses it; the AP3B2 row's prose argues for the term this
review proposes. I did not edit AP3B2: it is neither this gene's folder nor a donor in
any of its rows, and the campaign brief limits edits to those. Filed instead as
issue #3027.

Two actions retuned. GO:0006886 and GO:0016192 were ACCEPT while their own
reasons called them broad, harmless IEA parents; since ACCEPT means "retain as
representing the core function", both moved to KEEP_AS_NON_CORE with
root_cause: NO_FAILURE_NON_CORE. Counts are now 11 / 7 / 7 / 1 / 1.

One suggestion declined. The bot suggested adding qualifier: involved_in to the NEW
row for consistency with the seeded rows. Declined: the annotation-reviewer skill is
explicit that the gene-product-to-term relationship type "plays no role in review
reasoning, and any such value you encounter in a YAML row is likewise inert; never add,
edit, or argue from it." The seeded rows carry qualifiers because the GOA seeder copied
them, which is not a reason to author one. The NEW row's reason was instead reworded to
stop leaning on the word involved_in, and now argues the point directly.

Round 3: a denominator corrected. The WHOLLY_DARK gap first said "21 of the 26 rows
carry a WITH/FROM". Counted from the tsv, it is 20 — the six without are the five
ComplexPortal NAS rows and the ProtInc TAS row — and the 21st propagation_review in the
file belongs to the proposed NEW row. So the finding is 13 of 20, not 13 of 21, which
makes it slightly stronger. Also added to RESULTS.md §1: the pairwise identity matrix is
computed per ordered pair, so where several alignments score equally the two directions
can pick different ones and the matrix is very slightly asymmetric (largest
cell-vs-transpose difference 0.6 percentage points, none of it in the AP3M2 row the review
cites). The history record's details was rewritten to the final numbers; docs/history.md
freezes only target.slug and target.path, so details is editable in place.

2026-09-21 full-gene re-review

All 26 original assertions preserved; one old NEW GABA-secretion row withdrawn. Full PMID:15492041 attributes impaired evoked release at least partly to altered VGAT/vesicle biogenesis; AP3M2 performs cargo-sorting work, but this does not establish an additional GABA-release step. Exact GO:0061534 comparator queries on same-role human AP3B2/Q13367, AP3D1/O14617, AP3B1/Q92572 and AP3M1/Q9Y2T2 returned zero annotations (snapshot in audit JSON). The phenotype and mechanistic vesicle-budding function remain in references/core prose.

Actual PTHR10529 target leaf PTN002540363 descends from root PTN000055849 through AP3 node PTN002237676. TGN-loss PTN000242370 and transport-loss PTN002575694 are outside the target path. Full PMID:15051738 says “AP-3 does localize with clathrin, but to a lesser extent than does AP-1” and records 4% TGN labeling; full PMID:39705307 explicitly discusses clathrin-dependent AND independent functions. This overturns the previous categorical exclusions and broadens the standalone description. TGN is retained noncore, Golgi-vacuole transit remains compatible with an endosomal intermediate, and clathrin-associated complex membership is contextual. Exact clathrin-cargo MF scope remains uncertain.

GO:0035651 does not restrict binding to external ligands. Mouse donor Q8R2R9 IDA PMID:19010779 is real; full assay details remain inaccessible, so the old complex-subunit argument is replaced with UNDECIDED. The general AP-type coat term and cytoplasm/transport/axon-cytoplasm parents are valid core claims. GO:0016183 specifically names presynaptic clathrin-pit coating, so the source-supported endosomal-budding replacement is retained without asserting all other capacities absent.

Affinage body read: its mu-2 designation is wrong, and nonhuman findings are not human direct assays. The 9C5B structure is AP3M1, already established by local reproducible mapping; eight of nine cargo contacts are identical in AP3M2. Prior proposed ontology changes based on globally clathrin-free AP3 were withdrawn; non-endocytic signal-recognition vocabulary can be considered independently later. Focused clathrin/Golgi/source-binding report is queued.

Recovery review consistency follow-up (2026-09-22)

Restored readable identifiers in manual prose. Where applicable, reconciled AP3M2 reference notes with the retained contextual claim, removed unrelated PIK3C3 support from unresolved projections, separated PIK3C3 aspect-specific reasons, and documented the surviving/renamed ATG14 membrane term. Source assertion fields and verbatim quotations are unchanged.

Cytoplasm evidence and identifier correction (2026-09-23)

Restored ARBA accessions after an overbroad citation-formatting repair. GO:0005737
remains ACCEPT: recruitment of the neuronal p47/beta-NAP complex from brain cytosol
in PMID:9151686 supports a cytoplasmic coat-adaptor role. The delta redistribution
experiment is complex-level evidence from nonneuronal cells, not a direct AP3M2 assay.
The original ARBA condition-set analysis is preserved as a historical, version-limited
provenance concern and no longer confused with the biological action.