AP5M1 (human) — review notes

Working journal for the PAINT-backlog review of human AP5M1 / Q9H0R1, the
mu subunit of the fifth adaptor protein complex. Every assertion below carries
its provenance inline.

1. Identity check

AP5M1-uniprot.txt is ID AP5M1_HUMAN, AC Q9H0R1; (secondaries O95354,
Q6ZMD7, Q96DX3, Q9NVC5), DE RecName: Full=AP-5 complex subunit mu-1, 490 aa,
sequence version 2, OX NCBI_TaxID=9606. That is the protein the prompt names,
so no merged-accession substitution has happened. Gene names in the record:
GN Name=AP5M1; Synonyms=C14orf108, MUDENG, with protein-level aliases Mu5,
Mu-2-related death-inducing protein (MuD) and "Putative HIV-1
infection-related protein". The alias set matters: the literature is split
between papers that call this protein μ5/C14orf108 (membrane traffic) and papers
that call it MUDENG/MuD (cell death), and they barely cite one another.

Domain architecture, from the record itself:

UniProt's own summary is deliberately hedged: "As part of AP-5, a probable fifth
adaptor protein complex it may be involved in endosomal transport. According to
PubMed:18395520, it may play a role in cell death." Subcellular location:
Cytoplasm, cytosol {ECO:0000269|PubMed:22022230}, plus late endosome membrane
and lysosome membrane as ECO:0000305 (curator inference) with
Note=May cycle on and off membranes.

2. What AP-5 is

Hirst et al. 2011 (PMID:22022230) is the founding paper and is the source of
every experimental GOA row on this gene. Cached full text is available. The
relevant sequence of findings, in their own words:

The complex is actually a heterohexamer. Hirst 2013 (PMID:23825025,
abstract-only in cache) showed PMID:23825025,
PMID:23825025 and proposed the architecture:
PMID:23825025. Critically for this review, the
knockdown series included μ5: PMID:23825025

Mai et al. 2025 (PMID:40175557; abstract only in cache, full_text_available: false) solved it: PMID:40175557,
PMID:40175557,
PMID:40175557 and, the one molecular-activity statement anywhere in
the AP-5 literature, PMID:40175557,
PMID:40175557.

I checked the depositions behind that paper directly (RCSB REST,
data.rcsb.org/rest/v1/core/entry/…): 8YAB ("AP5 complex bound to
SPG11-SPG15") and 8YAH ("full length AP5 complex bound to SPG11-SPG15"). Both
are five-entity cryo-EM models. The μ5 entity in both maps to Q8BJ63, the
mouse orthologue, as does ζ (Q3U829), while β5 (Q2VPB7), σ5 (Q9NUS5) and
spatacsin (Q96JI7) are human — so the reconstitution is mixed-species, and no PDB
entry maps to human Q9H0R1 (consistent with the UniProt record carrying no DR PDB line). In 8YAH the μ5 chain is essentially fully modelled (unobserved only
at residue 1 and in five short stretches: 146–150, 188–190, 201–205, 285–290,
364–365). Mouse and human μ5 are 85.1% identical over the alignment (§4), so this
is a fair structural proxy, but it is a proxy, and it is the reason this review
does not treat the structure paper as direct evidence on the human protein.

3. What is known about mu-5 specifically, as opposed to about AP-5

This is the question that decides how the experimental rows are graded, so I
separated the two.

Evidence that perturbs or observes μ5 itself

Finding Source Grade
siRNA (SMARTpool + individual oligos) against C14orf108 → CIMPR and Vps26 in larger perinuclear puncta; swollen MVBs with emanating tubules PMID:22022230 IMP, μ5-specific
GFP-μ5 (after siRNA damping of endogenous protein) punctate, overlapping LAMP1 PMID:22022230 IDA, μ5-specific, tagged + perturbed background
μ5 partitions ~50:50 membrane/cytosol, absent from CCVs PMID:22022230 IDA, μ5-specific, endogenous
μ5 co-immunoprecipitates with σ5-GFP together with β5 PMID:22022230 IDA, μ5-specific
μ5 is one of the "six knockdowns" that trap CIMPR in early endosomes PMID:23825025 IMP, μ5-specific (abstract)
μ5 protein falls when AP5Z1 is lost or spastizin is depleted PMID:26085577 μ5 measured, not perturbed
Bi-allelic LoF in AP5M1 causes recessive macular dystrophy; anti-AP5M1 puncta in RPE PMID:40081374 human genetics + IDA-grade IF
MuD/AP5M1 is the μ subunit of AP5; TurboID proximity labelling PMID:41915273 abstract only

Evidence that is about AP-5 or the hexamer, with μ5 inferred

Because μ5 is destabilised when ζ is lost (Hirst 2015, quoted above), an AP5Z1
knockout is functionally an AP-5 knockout, and the CIMPR phenotype it produces is
the same one that μ5's own knockdown produces (Hirst 2011, Hirst 2013). That
concordance is what licenses giving AP5M1 the endosome-to-Golgi retrieval term;
it is not inferred from complex membership alone.

4. Does mu-5 have a working cargo-binding site? (the MHD question)

AP5M1 has a bona fide MHD by every domain resource (UniProt DOMAIN 206..476,
PROSITE PS51072, CDD cd09256 AP_MuD_MHD, Pfam PF00928). In AP-1/AP-2 the MHD is
the receptor for YxxΦ sorting signals, so "has an MHD" reads, wrongly, as "binds
cargo". Both Hirst papers say it does not, but neither names a residue:

I tested it rather than repeating it — AP5M1-bioinformatics/, written up in
RESULTS.md. Method in brief: derive the pocket from the solved complex (PDB
1BXX, μ2 second domain with the TGN38 DYQRLN peptide) by taking every μ2
residue within 4.5 Å of the peptide, after verifying the PDB author numbering
against Q96CW1 residue-by-residue; then map those 14 positions through a MAFFT
L-INS-i alignment of the five human μ subunits plus mouse and Arabidopsis μ5.

Result: AP5M1 retains 3 of the 14 (K240, I461, R465), against 11/14 for
AP4M1, 10/14 for AP1M1 and 5/14 for AP3M1. The three closest contacts to the
signal tyrosine — μ2 F174 (3.75 Å), D176 (2.43 Å, the shortest contact in the
interface) and W421 (3.06 Å) — are S210, S212 and A463 in μ5: no aromatic, no
acidic side chain, no bulk at any of the three. AP4M1 and AP1M1 keep all three;
even AP3M1, which also works without clathrin, keeps an aromatic and the
aspartate.

A third check arrived for free. The AP3M1 review merged onto main while this one
was in progress, and it asked the same question for μ3 using the same structure and
the same 4.5 Å cutoff, leaving a machine-readable results.json. Its script and its
sequence panel are different from mine, so the overlap is a replicate;
crosscheck_ap3m1.py compares the two and reports disagreements as well as
agreements. The structure-derived half is identical — same 14 positions, same
residues, same minimum distances to 0.01 Å — and the projected assignments agree
14/14 for AP1M1, 14/14 for AP3M1 and 12/14 for AP4M1, the two AP4M1 differences
being gaps in their alignment at mu2 V418 and I419, the only two contacts in neither
sub-site. Both runs independently score AP4M1 as retaining 11/14.

The obvious objection is that μ5 is simply the most divergent member (18.0%
identity to AP2M1), so any alignment-based count would look bad. Two controls
answer it. (i) The alignment reproduces a correspondence published independently
of this analysis: PMID:40081374 — my alignment puts AP5M1 Y313 in the same column as AP4M1
Y284, and finds a tyrosine there in all seven sequences. (ii) Mouse Ap5m1, 85.1%
identical to the human protein, gives the same residue at 13 of the 14 positions.

Conclusion carried into the review. μ5 has the MHD fold but not the
tyrosine-signal site, and no sorting motif of any kind has been identified for
AP-5. So: no cargo-binding or cargo-receptor MF is proposed for AP5M1, and the
"has an MHD therefore binds cargo" inference is explicitly blocked in the review
text. This is the mirror of the fold-name-propagates-into-GO error, checked in
the direction the brief asks for — I did not assert "fold without function"
without testing the residues.

5. The five IBAs — and why the AP3M1 argument does not transfer

All five IBA rows carry the identical WITH/FROM
PANTHER:PTN002699615|UniProtKB:Q9H0R1, reference GO_REF:0000033, dated
20190314. The committed PAINT slice
interpro/panther/PTHR16082/PTHR16082-paint.tsv resolves the node completely:

PTHR16082  PTN002699615  GO:0005764  C  IBD  seeds=UniProtKB:Q9H0R1  taxon:2759
PTHR16082  PTN002699615  GO:0005770  C  IBD  seeds=UniProtKB:Q9H0R1  taxon:2759
PTHR16082  PTN002699615  GO:0005829  C  IBD  seeds=UniProtKB:Q9H0R1  taxon:2759
PTHR16082  PTN002699615  GO:0030119  C  IBD  seeds=UniProtKB:Q9H0R1  taxon:2759
PTHR16082  PTN002699615  GO:0016197  P  IBD  seeds=UniProtKB:Q9H0R1  taxon:2759

Five IBDs, one node, one seed, no IRD or IKR anywhere in the family. The node is
placed at the eukaryotic root (taxon:2759), which matches the paper that made
the family: AP-5 subunits occur in all five supergroups but are repeatedly lost.
The family is small and clean — PTHR16082-entries.csv lists six reviewed
proteins, all named "AP-5 complex subunit mu-1"/"mu" (human, mouse, rat, macaque,
bovine, Arabidopsis), all in subfamily PTHR16082:SF2; family counters give
1889 member proteins across 3729 taxa, integrated as IPR039591.

The target is its own seed. This is the expected, valid configuration, not
circularity: AP5M1's own IDA/IMP annotations from Hirst 2011 are the descendant
evidence the PAINT curator used to place the IBD, and the IBA then adds the
claim that the localisation and process are ancestral to the family rather than
human-specific — a claim the 2011 phylogeny (five supergroups) directly
supports. None of these sources is marked CIRCULAR_OR_REDUNDANT.

Does the AP3M1 finding apply here? The merged genes/human/AP3M1/ review on
main found that the whole-family mu-subunit node carries a clathrin-cargo
adaptor term seeded only by AP-1/AP-2 mu subunits, and modified it to a
clathrin-free term. I checked whether AP5M1 inherits anything of the kind. It
does not, for three separate reasons:

  1. Different family. AP5M1's only PANTHER assignment is PTHR16082 ("AP-5
    COMPLEX SUBUNIT MU-1"), a family that contains AP-5 mu subunits and nothing
    else. It is not in the general AP mu-adaptin family, so no whole-family node
    reaches it. interpro/panther/PTHR10529/ is not present in this worktree and
    is not needed — AP5M1 is not in it.
  2. Different seeds. Every IBD above is seeded by AP5M1 itself, not by AP-1/
    AP-2 donors, so there is no cross-complex leakage to argue with.
  3. No clathrin term arrives. The only complex term that does arrive,
    GO:0030119 AP-type membrane coat adaptor complex, is not a
    clathrin-requiring term in its definition — it reads "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"
    (QuickGO /ontology/go/terms/GO:0030119/complete). AP-5 sits comfortably
    under it, and the child GO:0044599 AP-5 adaptor complex is explicitly
    clathrin-agnostic: "…it is not clear whether AP-5 forms clathrin coats in
    vivo    ". There is nothing to strip.

What is wrong with the GO:0030119 rows is only granularity: a term one level
more specific and exactly right — GO:0044599, whose ancestors include GO:0030119
(verified via the QuickGO ancestors endpoint) — has existed all along and is
already on the gene from ComplexPortal. Both GO:0030119 rows are therefore
MODIFY→GO:0044599, not REMOVE.

The other four IBAs are sound. Lysosome and late endosome are where the complex
works; is_active_in cytosol is true (μ5 really is ~50% soluble) but describes
the off-membrane reservoir rather than the site of action, so those rows are kept
as non-core rather than accepted as core. GO:0016197 endosomal transport is the
right breadth for a family-level assertion: the deeper step is known only in
human cells and only for a handful of cargo, and AP-5 is absent from most model
organisms, so pushing the node's term deeper would over-claim.

6. The three SubCell IEAs

GO:0005765 lysosomal membrane (SL-0157), GO:0005829 cytosol (SL-0091) and
GO:0031902 late endosome membrane (SL-0151) come from GO_REF:0000044, i.e. the
UniProt subcellular-location → GO mapping. The corresponding CC SUBCELLULAR LOCATION lines in AP5M1-uniprot.txt are Cytoplasm, cytosol {ECO:0000269|PubMed:22022230} (experimental) and late endosome / lysosome
membrane at {ECO:0000305|PubMed:22022230} (curator inference from the LAMP1
colocalisation and the membrane fractionation). The mapping is mechanically
faithful to the record; the two membrane terms inherit the 305-grade inference,
which is exactly the right strength — and the 2025 RPE work has since
corroborated the compartment independently in a second cell type
PMID:40081374, with co-staining for Rab7 and
TGN markers.

7. The MUDENG / cell-death literature

This is the whole of what the affinage record returned (§9), so it needs an
explicit verdict rather than silence.

Net: the cell-death phenotype is real in the sense of being reproducible in these
labs' hands, but its sign reverses between cell types (pro-death on
overexpression in Jurkat/HeLa/cervical carcinoma; anti-death on depletion in
astroglioma/GBM), and every positive result is an over- or under-expression
experiment with no mechanism tied to AP-5 membrane traffic. GOA carries no
apoptosis term for this gene and I am not proposing one; it goes in
knowledge_gaps and suggested_experiments instead. Note the direction of the
error to avoid: this is not grounds to doubt the trafficking annotations, and
it is not grounds to annotate apoptosis.

8. Human genetics

Kaminska et al. 2025 (PMID:40081374, full text): PMID:40081374. Three of the 19 families carry AP5M1 alleles —
c.97C>T (p.Arg33Ter), c.1166G>A (p.Trp389Ter) and the single missense in the
whole study, c.938A>G (p.Tyr313Cys). The paper notes the paralogous residue
(quoted in §4) is mutated in SPG50. This is the first disease association for
AP5M1 itself; AP5Z1 has long been SPG48. It corroborates that AP5M1 is
non-redundantly required in a lysosome-heavy tissue (RPE), which is the strongest
organismal support the endolysosomal annotations have.

No GO annotation is proposed from the disease association itself — a disease
phenotype is not a biological process the gene product executes.

9. What affinage missed

AP5M1-deep-research-affinage.md has self_evaluation_pairwise: win,
faith_pct: 100.0, 5 citations, and .affinage.log says "trust gates clear".
Its five cited findings (PMIDs 18395520, 23665015, 23909422, 27136675, 31427081)
are all real and all correctly summarised — and all five are MUDENG cell-death
papers. It returned nothing on AP-5. Its narrative ends: "Beyond its role in
apoptotic signaling, no clathrin- or adaptor-complex trafficking function has
been characterized for AP5M1 in the available corpus." That sentence is false
for the protein: the AP-5 characterisation (PMID:22022230) is the source of
nine of the twenty GOA rows and of the UniProt FUNCTION line, and there are at
least four further AP-5 papers (PMIDs 23825025, 26085577, 29381698, 40175557) and
a 2025 human-genetics paper (PMID:40081374).

This is the precision-not-recall failure the brief warns about, in its purest
form: the decisive papers are titled for the complex ("The fifth adaptor
protein complex") or for a partner ("…the hereditary spastic paraplegia
proteins SPG11 and SPG15"), never for the gene symbol, and the symbol-anchored
corpus is dominated by one small group's MUDENG series. Recovering them needed
independent searching on the complex name, the partners and the structure. The
record is therefore marked LOW_QUALITY in references[] — not because any
individual claim in it is wrong, but because its scope statement is wrong and
would mislead a curator who trusted it. (Europe PMC was intermittently returning
503s during this work; the PMIDs above were recovered through NCBI esearch and
targeted Europe PMC queries on "AP5M1", "AP-5 adaptor", "MUDENG".)

Its mechanism_profile grounding — GO:0005783 endoplasmic reticulum,
GO:0005739 mitochondrion, Reactome R-HSA-5357801 Programmed Cell Death — was
not imported, per §7.

10. Curation decisions

20 GOA rows over 11 distinct terms (16 cellular_component, 4 biological_process,
0 molecular_function), 0 NOT rows, 0 isoform-scoped rows. By evidence: 6 NAS,
5 IBA, 5 IDA, 3 IEA, 1 IMP; by source: 9 UniProt, 6 ComplexPortal, 5 GO_Central.
By reference: PMID:22022230 ×9, GO_REF:0000033 ×5, GO_REF:0000044 ×3,
PMID:40175557 ×3. Counts produced by
AP5M1-bioinformatics/check_goa_reconciliation.py and a one-off tally of the
GOA tsv, not by eye.

Term Codes Action
GO:0044599 AP-5 adaptor complex NAS ACCEPT (core complex membership)
GO:0030119 AP-type membrane coat adaptor complex IBA, IDA MODIFY → GO:0044599 (granularity)
GO:0005770 late endosome IBA, IDA, NAS ×2 ACCEPT
GO:0005764 lysosome IBA, IDA, NAS ACCEPT
GO:0031902 late endosome membrane IEA ACCEPT
GO:0005765 lysosomal membrane IEA ACCEPT
GO:0016020 membrane IDA MODIFY → GO:0031902 + GO:0005765
GO:0005829 cytosol IBA, IDA, IEA KEEP_AS_NON_CORE ×3
GO:0016197 endosomal transport IBA, IMP ACCEPT
GO:0016192 vesicle-mediated transport NAS MODIFY → GO:0016197
GO:0007040 lysosome organization NAS ACCEPT
GO:0042147 retrograde transport, endosome to Golgi — NEW (IMP)

One NEW row. No REMOVEs: nothing in this record is contradicted, and the two rows
that are merely uninformative (GO:0016020, GO:0016192) are true parents that
deepen cleanly rather than errors to delete.

11. Open questions carried into the review