AP5B1 (human, Q2VPB7) — review notes

Curation journal for the PAINT/affinage review wave. Provenance is inline as
[PMID:NNNNN "verbatim quote"]. Everything asserted here was checked against the cited
source; anything I could not check is flagged as such.

0. Inputs and identity check

• UniProt record fetched into AP5B1-uniprot.txt is ID AP5B1_HUMAN Reviewed; 878 AA.
  with primary accession Q2VPB7 — the expected accession, not a merged one. Secondary
  accessions A1L0S6, H6WUK2, Q0D2Q2, Q8N3J7, Q8WYH6.
• 878 aa, 93,949 Da. This matters historically: the sequence in the databases in 2011 was
  57 residues short at the N terminus, and Hirst et al. corrected it —
  PMID:22022230.
  The current record matches the corrected length exactly.
• Aliases that matter for literature searching: beta-5 / β5 (the name proposed in 2011),
  DKFZp761E198 (the clone name used throughout Hirst et al. 2011), ORF name PP1030.
  Searching Europe PMC for "beta5 adaptin" returns zero hits; the productive queries
  are AP5B1, "AP-5 adaptor", "adaptor protein complex 5", and the clone name.
• GOA: 24 rows, 9 distinct GO terms. By evidence code: IPI 12, NAS 6, IBA 2, IEA 1, HDA 1,
  IMP 1, IDA 1 (15 rows carry an experimental code; 15 rows carry a WITH/FROM). By aspect:
  MF 12, CC 7, BP 5. By assigning group: IntAct 8, UniProt 7, ComplexPortal 6, GO_Central 2,
  InterPro 1. Every one of these numbers was produced by a script over the GOA tsv, not read
  off by eye; the same harness reconciles each GOA row against exactly one review row on the
  key (GO id, evidence code, reference, normalised WITH/FROM) and it passes at 24/24 with one
  additional NEW row. The harness lives outside the repo tree, per the campaign rule on
  scratch files.
• Affinage record: self_evaluation_pairwise: win, faith_pct: 100.0, 7 citations, and
  .affinage.log says "trust gates clear". The narrative describes the right protein — the
  beta subunit of AP-5 — with no symbol collision of the AGT/AGXT kind. All 7 cited ids are
  numeric PMIDs; none is a bioRxiv PMID:bio_... id. All 7 resolved and were cached.

1. What the protein is

AP5B1 is the beta subunit of AP-5, the fifth and most recently discovered heterotetrameric
adaptor protein complex. The complex was defined in one paper: μ5 (C14orf108/AP5M1) was found
in a yeast two-hybrid screen to bind an uncharacterised protein, DKFZp761E198, whose closest
relatives are the beta adaptins —
PMID:22022230
— and the pair were renamed on that basis:
PMID:22022230.
The other two subunits, ζ (AP5Z1/KIAA0415) and σ5 (AP5S1/C20orf29), came from an independent
proteomics study published the year before PMID:20613862.

Architecture. β5 has the beta-adaptin layout — N-terminal alpha-solenoid trunk, then a
C-terminal appendage — but with two differences the 2011 paper called out:
PMID:22022230.
So it is more complete than β4 in its appendage and less flexible than β1/β2 in its hinge.
The cryo-EM structure of the whole assembly (PDB 8YAB/8YAH, chain B = AP5B1 1-878) resolves
helices from residue 10 to 629, i.e. the solenoid; UniProt annotates only one disordered
region in the whole protein, FT REGION 234..260.

No clathrin apparatus. This is the single most informative negative about β5, and I
re-tested it rather than repeating it — see AP5B1-bioinformatics/RESULTS.md. The 2011
statements were:
PMID:22022230 and
PMID:22022230.
My re-test on the current UniProt sequences and current Pfam boundaries:

• LLDLL and YQW are absent — confirmed.
• The clathrin box consensus L(phi)x(phi)[DE] matches once, LLRLE@841, inside the folded
  PF21590 domain. In AP1B1 and AP2B1 the same consensus matches once each (LLNLD@632,
  LLNLD@631) and both sit in the long unstructured hinge — the position where the motif
  actually works. So
  [file:human/AP5B1/AP5B1-bioinformatics/RESULTS.md "AP5B1 has no clathrin box in an accessible linker"].
  Caveat I would rather state than bury, since the same structural test decides the WDW
  question below: residue 841 is not modelled in 8YAB chain B (span 7-631), so the
  "buried in a folded domain" call rests on the Pfam assignment alone, and an unmodelled
  region is often a flexible one. The conclusion does not depend on it — the AP1B1/AP2B1
  boxes are not merely unstructured but sit in a linker three times longer, and AP-5's
  clathrin independence is a biochemical result.
• The trunk-to-appendage linker is 62 aa in AP5B1 against 187 in AP1B1 and 178 in AP2B1;
  [file:human/AP5B1/AP5B1-bioinformatics/RESULTS.md "AP5B1's is the shortest, and by a factor of ~3 against the two clathrin-dependent beta subunits."]
  This confirms the "no long hinge" claim quantitatively.
• The WDW at 223-225 is in the solenoid, and the 2011 reading holds. My first pass got
  this wrong and the PR review caught it, so the correction is recorded here rather than
  quietly overwritten. I had read the absence of a covering Pfam domain and of a covering
  UniProt feature at residue 223 as evidence that the motif sits outside the solenoid in an
  unstructured inter-domain segment, which would have made Hirst et al.'s structural
  dismissal of it wrong. That was an over-reading of an absence: neither database annotates
  inter-helical loops. Adding a PDBe polymer-coverage check to the script settles it —
  residues 223-229 are modelled in 8YAB chain B, whose modelled span is 7-631, so the
  tripeptide is inside the cryo-EM-resolved trunk, bracketed by Helix 182-203 and the
  MobiDB-lite Region 234-260 (Disordered) and flanked by unmodelled stretches 213-222 and
  230-260. It is a short ordered island in a poorly ordered inter-helical loop of the
  solenoid, not an exposed hinge. The apparent inter-domain position is a Pfam
  family-model coverage gap (105-262) and nothing more. Independently of where it sits, it
  is not a clathrin motif:
  [file:human/AP5B1/AP5B1-bioinformatics/RESULTS.md "the validated W-box consensus is PWxxW, and PWxxW is absent from AP5B1"],
  and the cell biology never depended on the sequence argument:
  PMID:22022230.

Family placement. UniProt puts AP5B1 in DR PANTHER; PTHR34033; AP-5 COMPLEX SUBUNIT BETA-1; 1.
and DR PANTHER; PTHR34033:SF1; AP-5 COMPLEX SUBUNIT BETA-1; 1. — not in the beta-adaptin
family. AP1B1, AP2B1, AP3B1 and AP4B1 are all PTHR11134 ("ADAPTOR COMPLEX SUBUNIT BETA FAMILY
MEMBER", 17,626 proteins). PTHR34033 has 1,227 proteins, one subfamily, and five reviewed
members: the human, mouse, rat, bovine and Xenopus AP5B1 orthologues
(interpro/panther/PTHR34033/PTHR34033-entries.csv). The practical consequence for this
review is that the PAINT route by which AP-1/2/3/4 beta-subunit terms could leak onto AP5B1
does not exist —
[file:human/AP5B1/AP5B1-bioinformatics/RESULTS.md "PAINT therefore cannot leak AP-1/2/3/4 beta-subunit terms onto AP5B1 through the family tree."]
The two IBAs on this gene are therefore not a family-transfer risk; they are something else
entirely, discussed in §4.

2. The complex(es): AP-5 core and the AP-5/SPG11/SPG15 hexamer

Two distinct assemblies, and GO represents only one of them.

AP-5 core (tetramer) = ζ (AP5Z1) + β5 (AP5B1) + μ5 (AP5M1) + σ5 (AP5S1). GO term
GO:0044599 AP-5 adaptor complex, definition "An AP-type membrane coat adaptor complex that
in humans consists of beta5, zeta, mu5 and sigma5 subunits and is found associated with
membranes in the endosomes; it is not clear whether AP-5 forms clathrin coats in vivo."
ComplexPortal CPX-5181, 1:1:1:1, which is what the UniProt record cites
(DR ComplexPortal; CPX-5181; AP-5 Adaptor complex.).
PMID:26085577

AP-5/SPG11/SPG15 (hexamer) = AP-5 core + spatacsin (SPG11) + spastizin (ZFYVE26/SPG15),
stoichiometric:
PMID:23825025.
This is the form the complex is thought to take natively —
PMID:29381698.
The division of labour proposed in 2013:
PMID:23825025
The 2025 cryo-EM structure supports it:
PMID:40175557 and
PMID:40175557

GO has no term for the hexamer. I checked: QuickGO ontology search for "spatacsin" and
for "SPG11" returns nothing; OLS search of GO for "spastizin spatacsin complex" returns
nothing; GO:0044599 has no children. ComplexPortal, by contrast, does curate it —
CPX-20045 "AP5-Spastizin-spatacsin complex" (participants CPX-5181 + CPX-26503, cited to
PMID:40175557, PMID:25365221, PMID:23825025) and CPX-26503 "Spastizin-spatacsin complex" —
and neither carries a GO cross-reference. That asymmetry is a genuine ontology gap, recorded
as such in the review's knowledge_gaps and proposed_new_terms. It is also why the four
SPG11 IPI rows and the one SPG15 IPI row cannot be replaced by the CC annotation that would
actually capture them. Because CPX-20045 is asserted as a machine-readable
skos:exactMatch and was the only load-bearing identifier in the review with no other
committed source, the fetch is now scripted and its output committed:
AP5B1-bioinformatics/complexportal_provenance.py, complexportal.json,
complexportal-output.txt. The script asserts CPX-20045 still contains Q2VPB7 and prints a
retirement notice if a GO cross-reference ever appears, which is how a later curator learns
the proposed term is no longer needed. (Incidental finding from the same fetch, not built on
here: ComplexPortal also lists CPX-13475 "CDC123:AP5B1" and CPX-19795, both putting CDC123
(O75794) with AP5B1. CDC123 appears in no AP-5 paper I read and in no GOA row for this gene.)

Assembly evidence specific to AP5B1 (i.e. not just "AP-5 does X"):

• β5 binds μ5 directly, and the site is mapped:
  PMID:22022230
  PMID:22022230
  This is the canonical beta-mu arrangement of AP-1/AP-2 reproduced in AP-5.
• β5 is required for μ5 stability, the classic beta-subunit behaviour:
  PMID:22022230.
• β5 was itself used as a bait in the study that found ζ and σ5:
  PMID:20613862,
  PMID:20613862.
  So the AP5B1 IPI rows attributed to PMID:20613862 are bait-side as well as prey-side.
• AP5B1 is named as a component of the assembled complex a decade later:
  PMID:33464297.
• A measured negative worth recording: β5 was used as bait in a yeast two-hybrid library
  screen looking for a second large subunit and found none —
  PMID:22022230.
  ζ was found by proteomics, not by two-hybrid.

3. Function

Endosome-to-Golgi retrieval. The AP5B1-specific evidence is the knockdown phenotype. β5
siRNA phenocopies μ5 siRNA: the cation-independent mannose 6-phosphate receptor and the
retromer subunit Vps26 accumulate in enlarged perinuclear puncta —
PMID:22022230,
PMID:22022230,
and at the complex level
PMID:22022230.
The direction of the defect was not resolved in 2011 and the 2011 GO annotation reflects that
(GO:0016197 endosomal transport, IMP). It was resolved in 2018 by CRISPR knockout of AP5Z1
plus spatial proteomics:
PMID:29381698,
PMID:29381698.
Six knockdowns (four AP-5 subunits plus SPG11 and SPG15) give the same CIMPR phenotype:
PMID:23825025.
Together this is why GO:0016197 endosomal transport (IMP) should become
GO:0042147 retrograde transport, endosome to Golgi — a descendant of GO:0016197 (verified
via the QuickGO ancestors endpoint), so this is a refinement, not a change of claim.

The cargo question is open, and the field says so. AP-5's own cargo-binding activity has
never been demonstrated:
PMID:29381698,
and even after identifying the pathway the cargo link is a proposal routed through a partner
subunit, not through AP-5:
PMID:29381698.
This is why I did not add GO:0140312 cargo adaptor activity as a NEW row or as
contributes_to_molecular_function: its definition requires bridging membrane, cargo receptor
and deformation machinery, and the cargo-receptor half of that has not been shown for AP-5.
It is recorded as an MF knowledge gap instead.

Recruitment. Coincidence detection of PI3P and Rag GTPases, with PI3P read by the SPG15
FYVE domain, so the membrane-binding determinant is on a partner rather than on β5:
PMID:33464297,
PMID:33464297

Membrane remodelling and lysosome reformation. The 2025 structural work adds an in vitro
activity for the assembled hexamer:
PMID:40175557,
PMID:40175557.
Note: that paper's cache entry is abstract-only (full_text_available: false), so I have not
been able to see which subunit contributes the curvature-sensing surface — and the abstract
also says the authors solved SPG11-SPG15 on its own, so I cannot even establish that AP-5 is
required for the activity. I first recorded this as a complex-level
contributes_to_molecular_function: GO:0140090, and dropped it after the PR review pointed
out the obvious: that slot would have asserted, in a field downstream tooling believes,
exactly what the accompanying knowledge gap declares undetermined, and would have attached it
to GO:0044599, the tetramer, when the measurement was on the hexamer. The gap is the honest
record; the slot is left empty.
The loss-of-function counterpart is a storage phenotype:
PMID:26085577

Localisation. Three independent lines, spanning 18 years:
- Placental lysosomal-membrane proteomics found it before anyone knew what it was —
PMID:17897319,
PMID:17897319.
- The complex colocalises on a late endosomal/lysosomal compartment —
PMID:23825025;
in 2011 the compartment was identified through μ5, because tagged β5 could not be localised —
PMID:22022230.
- AP5B1 protein itself was finally imaged in 2025, in human RPE, with an anti-AP5B1
antibody — PMID:40081374.
This is the first direct, subunit-specific localisation of AP5B1 and it agrees with the
complex-level NAS annotations from ComplexPortal.

Human genetics. AP5B1 is a macular dystrophy gene, established in 2025-2026 across three
independent cohorts. This is new since the GOA snapshot and none of it is annotated:
PMID:40081374,
PMID:41830174,
PMID:42568187.
The AP5B1 alleles in the 2025 cohort are truncating —
PMID:40081374 —
and p.Leu785Pro is a founder allele —
PMID:42568187.
Note that Leu785 falls inside PF21590 (781-873), the C-terminal appendage domain, per the
bioinformatics run. Unlike AP5Z1 (spastic paraplegia SPG48), AP5B1 has not been reported as an
HSP gene; the phenotype reported for AP5B1 is retinal.

A measured negative. The paper that discovered AP5Z1 came out of a DNA-repair RNAi screen,
and three of the five complex members scored in it — but β5 did not:
PMID:20613862
(the "did no have" typo is in the source and is reproduced verbatim). There is no DNA-repair
annotation on AP5B1 in GOA, so there is nothing to remove; recorded so that a future reviewer
reading the title of PMID:20613862 does not add one.

Not annotated, and deliberately so. HIV-2 Gag particle release depends on AP-5 while
HIV-1's does not — PMID:27392064 —
but the siRNA used there targets AP5M1 ("AP-5μ ON-TARGETplus AP5M1 J-015523-09"), not AP5B1,
and viral hijacking of a trafficking route is not a function of the host gene. No row.

4. The two IBAs and the one IEA

Fetched slice: just fetch-panther-paint PTHR34033 →
interpro/panther/PTHR34033/PTHR34033-paint.tsv, which contains exactly two IBD rows:

PTHR34033  PTN002145545  GO:0030119  C  IBD  false  UniProtKB:Q2VPB7  taxon:2759  20170228
PTHR34033  PTN002145545  GO:0016197  P  IBD  false  UniProtKB:Q2VPB7  taxon:2759  20170228

Reading this carefully matters. There is one node in the whole family, PTN002145545,
placed at taxon:2759 (Eukaryota) — the root of the family — and its only seed is
UniProtKB:Q2VPB7, i.e. AP5B1 itself. The two terms it carries are precisely the two
experimental annotations AP5B1 had in 2017: the IDA for GO:0030119 and the IMP for GO:0016197,
both from PMID:22022230. The GOA WITH/FROM for both IBA rows is
PANTHER:PTN002145545|UniProtKB:Q2VPB7, which is exactly what that node structure predicts.

So these are the textbook self-referential IBA case that CLAUDE.md and the skill both call out:
the gene's own experimental annotation is one of the descendant evidences the PAINT curator
used to place the IBD, so the gene legitimately appears among the sources of the IBA it
receives. That is not circular and the short donor list is not weak. What the IBA adds over
the IDA/IMP is a phylogenetic statement: the curator judged this function to be ancestral to
the whole eukaryotic AP5B1 clade rather than lineage-specific — a judgement that the
comparative genomics in the same paper supports
(PMID:22022230,
PMID:22022230).
root_cause: NO_FAILURE_CORE for both, source_status: SUPPORTS_TRANSFER for both entries.
Human AP5B1 is trivially inside the inheriting clade (it is the seed), there is no IRD/IKR
anywhere in the family, and the family index shows the four other reviewed members are all
full-length (876-883 aa) orthologues with no sign of pseudogenisation.

The IEA row (GO:0016197, GO_REF:0000002, WITH/FROM InterPro:IPR038741) traces to an
InterPro2GO mapping I verified at the source: the InterPro API returns for IPR038741 the name
"AP-5 complex subunit beta-1", type family, a single GO mapping GO:0016197 endosomal transport, and the description "AP5B1 is part of the AP-5 adaptor protein complex, which may
be involved in endosomal transport". IPR038741 is the AP5B1-specific family signature (UniProt
carries it as DR InterPro; IPR038741; AP5B1.), so the mapping is family-specific, not a
promiscuous pan-adaptin mapping, and it is correct. root_cause: NO_FAILURE_CORE.

5. Partner resolution for the 12 IPI rows

All eight distinct accessions resolved via the UniProt REST API (size > 1, Swiss-Prot status
recorded):

Reference projection test (QuickGO by reference=, paginated fully, entities counted not
annotations):

The two proteome-scale screens are what they look like: 937- and 2,196-entity projections
(PMID:25416956,
PMID:33961781).
CAMK2B and NIF3L1 come only from those two and have no functional follow-up anywhere in the
AP-5 literature, so both go to MARK_AS_OVER_ANNOTATED. Note the trap: CAMK2B appears in both
screens and the UniProt IntAct block reads
Q2VPB7; Q13554: CAMK2B; NbExp=4; IntAct=EBI-5917279, EBI-1058722; — but an NbExp count is
experiments, not independent studies, and two high-throughput methods agreeing is still two
high-throughput methods. By contrast
Q2VPB7; Q96JI7: SPG11; NbExp=8; IntAct=EBI-5917279, EBI-2822128; is supported by targeted
work across four papers.

The lysosomal proteomics reference (242 entities) is a broad projection too, but unlike the
interactome screens its claim is independently corroborated for this protein by three later,
targeted studies (Hirst 2013, Hirst 2021, Kaminska 2025), so it is accepted rather than
downgraded.

6. Decisions taken (summary; the reasoning per row is in the YAML)

• GO:0030119 IDA (PMID:22022230) → MODIFY to GO:0044599 AP-5 adaptor complex. The
  cited paper is the one that defined AP-5 and named β5 as its subunit; the specific term
  now exists and is already on the gene by NAS. The strongest evidence should carry the
  specific term.
• GO:0030119 IBA → ACCEPT. The IBD mirrors the 2017 state of the seed; a family-level
  claim at the parent is appropriately conservative, not a granularity failure.
• GO:0016197 IMP (PMID:22022230) → MODIFY to GO:0042147 retrograde transport, endosome to Golgi (a verified descendant), with PMID:29381698 as additional reference for
  the direction.
• GO:0016197 IBA and IEA → ACCEPT (both correct, both appropriately general).
• GO:0016192 NAS → MODIFY to GO:0042147. It is the grandparent of a term already on
  the gene, so it conveys nothing as it stands.
• Nine GO:0005515 IPI rows whose partner is an AP-5 core subunit or a hexamer partner →
  MODIFY to GO:0005198 structural molecule activity. Justification is not "it binds
  something": β5 binds μ5 through a mapped N-terminal site in the canonical beta-mu
  arrangement, and knocking β5 down destabilises μ5 — that is a structural-integrity
  contribution, which is what GO:0005198 means. The annotation that would really capture the
  SPG11/SPG15 rows is CC membership of the hexamer, and that term does not exist.
• Three GO:0005515 IPI rows with CAMK2B/NIF3L1 → MARK_AS_OVER_ANNOTATED.
• The five CC/BP NAS rows and the HDA row → ACCEPT.
• One NEW row: GO:0001895 retina homeostasis, IMP, PMID:40081374. Three independent
  2025-2026 cohorts, bi-allelic LoF and founder missense alleles, plus subunit-specific RPE
  localisation. Flagged in its own reason as a tissue-level, non-core consequence rather than
  a molecular core function.

Final tally: 12 MODIFY, 9 ACCEPT, 3 MARK_AS_OVER_ANNOTATED, 1 NEW. No REMOVE and no
UNDECIDED: nothing in GOA for this gene is contradicted by the literature, and every cited
reference was readable at least in abstract.

Three validation warnings remain, all deliberate. Two are "inconsistent review actions"
on GO:0016197 and GO:0030119, where the IBA/IEA rows are ACCEPTed at the general term while
the human experimental row is refined. That divergence is the point rather than an oversight:
the direction of transport and the four-subunit composition are human results, and pushing
either down a level on a eukaryote-root PAINT node or on an InterPro family signature would
propagate a human-specific finding to 1,227 proteins across 2,973 taxa. Each of those four
review.reason fields states the divergence explicitly. The third warning is that no
supporting_text cites the affinage file; that is the campaign rule (affinage is a lead, not
evidence) and is recorded in the affinage reference_review.

7. What affinage missed, and what it got right

Got right: the mechanistic narrative is accurate and correctly scoped to AP-5 and to this
subunit; the retromer-backup framing, the Rag/PI3P coincidence detection, the LSD phenotype
and the 2025 macular dystrophy paper are all real and correctly cited. The mechanism_profile
GO grounding was not imported (per the brief) and would have been partly wrong if it had been:
it proposes GO:0060090 molecular adaptor activity as the molecular activity, which overstates
what has been shown, given that AP-5's cargo interaction is explicitly unknown.

Missed — found by independent Europe PMC searching:

1. PMID:42568187 (2026), the largest AP5B1 disease cohort to date, 22 individuals from 20
   families, and the paper that establishes p.Leu785Pro as a founder allele. Affinage's newest
   citation is PMID:40081374 (April 2025).
2. PMID:41830174 (2026), the independent replication cohort (AP5Z1 and AP5B1 retinal
   degeneration).
3. PMID:20613862 is in GOA and supplies four of the twelve IPI rows, and it is the study
   that found ζ and σ5 with β5 as one of the baits — affinage does not cite it at all. Its
   title is about a DNA repair screen, which is exactly the "titled for something else" failure
   mode the brief warns about.
4. PMID:22022230, the paper that defines the complex and contains every AP5B1-specific
   experiment, is also not in affinage's citation list (it cites the 2013 follow-up instead).
   Both 3 and 4 were already seeded into the review by GOA, so nothing was lost, but it is
   a reminder that affinage's gates measure precision, not recall.
5. PMID:17897319, the source of the HDA lysosomal-membrane row.

Cross-check on affinage's claims that I could verify independently: PMID:29381698,
PMID:26085577, PMID:33464297, PMID:27392064 and PMID:40081374 all resolve to the papers
affinage describes and say what affinage says they say. Reference reviews recorded accordingly.

8. Open questions carried into knowledge_gaps

• No demonstrated molecular activity for AP-5 or for β5: no cargo sorting motif recognised, no
  cargo binding shown. Pharos; Q2VPB7; Tdark.
• No GO CC term for the AP-5/SPG11/SPG15 hexamer even though ComplexPortal curates it
  (CPX-20045), so five IPI rows have nowhere accurate to land.
• Which subunit of the hexamer provides the curvature-sensing surface in PMID:40175557 is not
  determinable from the abstract, and that paper's cache entry has no full text.
• Whether β5's appendage — which uniquely among the AP-5-family-adjacent beta subunits retains
  both platform and sandwich subdomains — binds accessory proteins the way the AP-1/AP-2 ears
  do. No ligand has been reported for it.