Human Protein Atlas: Primary Cilium Life Cycle Module

Bottom line: The project builds a module for the life cycle of the
primary cilium, from licensing of the mother centriole to resorption before
mitosis, and uses the Human Protein Atlas (HPA) cilium atlas as localization
evidence for its members. All 60 module members now have gene reviews: 54 new
reviews (2,996 annotations, each with falcon deep research) plus 6 existing
ones. The module,
modules/primary_cilium_life_cycle.yaml,
was revised from those reviews and now has 8 stages, 34 role annotons and 60
human proteins. It passes schema and module validation.

The project started as a general re-evaluation of HPA subcellular evidence in
GO. That analysis is kept on a sub-page:
HPA subcellular evidence re-evaluation.

Why combine HPA with a cilium module

The HPA primary-cilium atlas
(Hansen et al. 2025, Cell, PMID:41005307)
mapped 715 proteins to four sub-ciliary classes (cilium, ciliary tip,
transition zone, basal body) in three cell lines. About 85% of its calls are
graded Approved or Uncertain, so they never reach GOA (see the sub-page).
On their own, those calls cannot tell a real ciliary protein from antibody
noise. A module supplies the missing context: if a protein has a defined role
at a defined life-cycle stage, its atlas call can be judged against that
role. Going the other way, the atlas tests whether each module member is
seen where the module puts it.

Module structure

Each stage cites primary literature checked against PubMed (distal
appendages PMID:23348840, TTBK2 PMID:23141541, EHD1/EHD3 PMID:25686250,
transition zone PMID:21725307, dynein-2 PMID:31451806, Aurora A–HDAC6
PMID:17604723, KIF2A PMID:25660017, and others in the YAML).

Boundary. The module covers primary (non-motile) cilia. Motile cilia,
centriole duplication, ciliary signaling pathways (Hedgehog is
modules/hedgehog_signaling.yaml) and the
ciliary pocket are outside it.

Ontology gaps found while building it (recorded in the module notes; the one genuine biological unknown, the molecular activity of the structural scaffold proteins, is the module's knowledge_gaps entry):

• GO has no term for regulation of cilium length and no cytoplasmic dynein-2
  complex term.
• Distal appendage scaffolds, transition zone barrier proteins, IFT-B scaffold
  subunits and microtubule depolymerases have no molecular-function term. These
  annotons carry a free-text function only.

Separately, primary cilium (GO:0072372) is obsolete. This is a term choice,
not a gap: the module uses non-motile cilium (GO:0097730) as context.

HPA evidence for module members

From cilium_life_cycle/member_evidence.md,
counting cilium, tip, transition zone and basal body calls for the 68 candidate
members:

The middle row is the key finding. Canonical ciliogenesis proteins
(CEP83, CEP164, CEP290, IFT52, KIF3A) carry only Approved or Uncertain grades.
For core ciliary proteins, HPA's grade says little about whether the location
is real. Module membership and literature are the better guide. The
"no call" row is mostly transmembrane transition zone proteins and
low-abundance motors, which antibody IF often misses.

Files

• Module: modules/primary_cilium_life_cycle.yaml
  (generated by scripts/build_cilium_module.py
  from a curated spec; UniProt names are read from the candidate table)
• Candidates: cilium_life_cycle/candidate_members.yaml
  (accessions from the UniProt REST API)
• Member evidence: scripts/cilium_members.py produces
  cilium_life_cycle/member_evidence.tsv and .md
• HPA cilia analysis: scripts/hpa_cilia.py and data/hpa_cilia_*
• HPA data snapshot (v25): data/subcellular_location.tsv

uv run python projects/HUMAN_PROTEIN_ATLAS/scripts/cilium_members.py
uv run python projects/HUMAN_PROTEIN_ATLAS/scripts/build_cilium_module.py
uv run linkml-validate -s src/ai_gene_review/schema/gene_review.yaml -C ModuleReview modules/primary_cilium_life_cycle.yaml
uv run python -m ai_gene_review.validation.module_validator modules/primary_cilium_life_cycle.yaml

STATUS

Last updated: 2026-10-03

Module

• [x] Define boundary and stages
• [x] Fetch UniProt accessions for candidates (UniProt REST)
• [x] Draft modules/primary_cilium_life_cycle.yaml (schema and module validation pass)
• [x] Review all 60 members and revise the module from the reviews (2026-10-03; see notes)
• [ ] Module deep research (just module-deep-research-falcon modules/primary_cilium_life_cycle.yaml) and reconcile with the draft
• [ ] Replace GENE_PRODUCT selectors with PANTHER families and PAINT nodes where they can be verified from local data
• [ ] Add HPA atlas localization evidence (PMID:41005307) to annotons where the member's review accepts it
• [ ] Decide whether the 8 candidates left out of the module belong in a stage: OFD1, KIF17, MPHOSPH9, PACSIN1, RAB11A, AHI1, LZTFL1 (BBSome regulator) and BBS1 (covered by modules/bbsome.yaml)
• [ ] Propose GO terms: regulation of cilium length; cytoplasmic dynein-2 complex; ATP-dependent microtubule depolymerase activity (KIF2A, KIF24)
• [ ] Expert check of the flagged REMOVE decisions on experimental annotations (see 2026-10-03 notes)
• [ ] Decide whether the IFT and transition-zone stages should reference modules/intraflagellar_transport.yaml and modules/ciliary_transition_zone.yaml (added on main in #4425)
• [ ] Resolve the 14 UNDECIDED annotations across the new reviews

Member gene reviews (all done 2026-10-03)

Already reviewed: human/TMEM67, human/CC2D2A, human/B9D1, human/ARL13B, human/HDAC6, human/PLK1 (plus candidates human/AHI1, human/BBS1, human/LZTFL1)

Stage 1, licensing:
- [x] human/CEP164
- [x] human/TTBK2
- [x] human/CEP83
- [x] human/SCLT1
- [x] human/CEP89
- [x] human/FBF1
- [x] human/CCP110
- [x] human/CEP97
- [x] human/KIF24

Stage 2, ciliary vesicle:
- [x] human/MYO5A
- [x] human/EHD1
- [x] human/EHD3
- [x] human/RAB3IP
- [x] human/RAB8A

Stage 3, transition zone:
- [x] human/CEP290
- [x] human/MKS1
- [x] human/TCTN1
- [x] human/TCTN2
- [x] human/B9D2
- [x] human/TMEM216
- [x] human/TMEM231
- [x] human/NPHP1
- [x] human/NPHP4
- [x] human/RPGRIP1L

Stage 4, IFT:
- [x] human/IFT88
- [x] human/IFT52
- [x] human/IFT81
- [x] human/IFT74
- [x] human/IFT172
- [x] human/IFT140
- [x] human/IFT122
- [x] human/WDR35
- [x] human/TTC21B
- [x] human/KIF3A
- [x] human/KIF3B
- [x] human/KIFAP3
- [x] human/DYNC2H1
- [x] human/DYNC2LI1
- [x] human/DYNC2I1
- [x] human/DYNC2I2

Stage 5, membrane composition:
- [x] human/ARL3
- [x] human/INPP5E
- [x] human/TULP3
- [x] human/RP2
- [x] human/PDE6D
- [x] human/UNC119B

Stage 6, length control:
- [x] human/CILK1
- [x] human/MAK
- [x] human/KIF7

Stage 7, disassembly:
- [x] human/AURKA
- [x] human/NEDD9
- [x] human/CIMAP3
- [x] human/KIF2A
- [x] human/DYNLT1

For each review, record how the HPA cilium call (grade, sub-ciliary class)
compares with the role the module assigns.

NOTES

2026-10-09 (merge with main)

• While this branch was open, main merged its own reviews of 14 module
  members: AURKA, B9D2 and KIF7 (#4350, #4358, #4404), and CEP290, CILK1,
  DYNC2H1, IFT140, IFT88, KIF3A, MKS1, NPHP1, NPHP4, RPGRIP1L and TCTN1
  (#4425). Main's reviewed versions are kept for those genes. This branch
  contributes the other 40 member reviews. The module still passes its
  checks against main's versions.

• 4425 also added modules/intraflagellar_transport.yaml and

  
  modules/ciliary_transition_zone.yaml, which cover the same ground as this
  module's IFT and transition-zone stages (and MAK/CILK1 regulation). Open
  question: whether those stages here should reference the dedicated modules
  instead of restating them.

2026-10-03 (member reviews and module revision)

• Reviewed all 54 unreviewed members with 10 parallel agents grouped by
  stage. Each gene got falcon deep research (needs --timeout 2400; the
  default 600 s times out and the perplexity-lite fallback is unavailable in
  this environment), cached publications, a notes file with an "HPA cilium
  atlas vs module role" section, a full review and a history record.
• 2,996 annotations: ACCEPT 1,404, KEEP_AS_NON_CORE 768, REMOVE 574 (mostly
  bare protein binding), MODIFY 144, MARK_AS_OVER_ANNOTATED 63, NEW 29,
  UNDECIDED 14. All 54 reviews validate with 0 errors; 53 have no warnings.
  MYO5A keeps one intentional consistency warning. The ARL3 falcon report
  arrived after its review was written and was then folded in; it changed no
  actions.
• Module changes driven by the reviews:
• Stage 2 split: RAB3IP/RAB8A act after the ciliary vesicle forms (RAB8 is
  activated only after vesicle assembly), so they are now a separate
  membrane-extension stage. MYO5A acts upstream of vesicle formation and now
  carries microfilament motor activity. EHD1/EHD3 carry ATP hydrolysis
  activity (EHD proteins bind ATP, not GTP).
• Licensing: TTBK2 triggers cap removal through MPHOSPH9/CEP83 (CP110 is
  not a known substrate); KIF24 is the cap recruiter, split from the
  CP110-CEP97 cap; the cap is located at the centriole. CEP89 also positions
  NCS1 for vesicle capture; FBF1 acts after licensing as an IFT/membrane
  gate.
• Transition zone: RPGRIP1L is now its own upstream annoton; NPHP4 is the
  adaptor hub of the NPHP module (protein-macromolecule adaptor activity);
  NPHP1 is downstream. CEP290 is not an MKS subunit and also acts at
  centriolar satellites.
• IFT: cargo binding (tubulin) is shown only for IFT81-IFT74, so they are
  split from the IFT-B scaffold subunits. IFT-A gains protein localization
  to cilium and protein carrier activity (with TULP3). KIFAP3 has no motor
  domain and is now a separate kinesin-binding annoton. Dynein-2 uses
  cytoplasmic dynein complex.
• Membrane: ARL3 uses G protein activity (its intrinsic GTP hydrolysis is
  negligible); PDE6D (farnesylated protein binding) and UNC119B (lipid
  binding) are separate carriers; TULP3 is a protein-macromolecule adaptor.
• Length control uses negative regulation of non-motile cilium assembly;
  KIF7 is a non-motile plus-end binder that slows microtubule growth. MAK's
  length evidence is mostly photoreceptor.
• Disassembly: NEDD9 and CIMAP3 are separate kinase-activator annotons;
  TTBK2 restrains KIF2A during growth (PMID:39930500).
• Decisions to check: two REMOVE actions overrule experimental annotations
  on biological grounds, which CLAUDE.md allows only for genuinely
  contradicted claims. (1) IFT-A membership (IDA, PMID:26980730, abstract
  only) on five IFT-B subunits: IFT88, IFT52, IFT81, IFT74, IFT172. All five
  are canonical IFT-B subunits, and the same paper supports their IFT-B
  annotations. (2) The ARL3 NOT-located-in-cilium row (PMID:17646400). Checked
  against the full text on 2026-10-04: the NOT rests on overexpressed EGFP-Arl
  constructs not targeting primary cilia in an RPE1 screen (Fig. S2 A). That
  is a negative enrichment result, not evidence that the function lies outside
  the cilium. ARL3-GTP is generated and acts inside the cilium (ARL13B, its
  GEF, is confined there; PMID:26551564), so REMOVE stands. The reason now
  quotes the paper directly.
• HPA: the HPA atlas paper (PMID:41005307) is cached as abstract only, so the
  reviews used the grades in member_evidence.md. Reviewers repeatedly found
  canonical ciliary proteins with weak or missing HPA calls: CCP110 is seen
  only at centriolar satellites, INPP5E only at the Golgi, and DYNC2H1,
  TMEM216, WDR35 and IFT172 have no call. This supports judging atlas calls
  in module context.

2026-10-03 (pivot)

• Pivoted the project from general HPA evidence re-evaluation to a primary
  cilium life-cycle module, with the HPA cilium atlas as evidence. The earlier
  work moved to HUMAN_PROTEIN_ATLAS/hpa-evidence-reevaluation.md; its
  worklists are paused, not abandoned.
• Kept the HUMAN_PROTEIN_ATLAS slug to avoid renaming churn.
• Built the draft module with a generator script so UniProt names come from
  the API-derived table, not memory. GO ids were checked in OLS, and 21 PMIDs
  were checked in PubMed. Three PMIDs that I first had in mind resolved to
  unrelated papers and were dropped.
• The module validator rejects complex terms (IFT-A, IFT-B, kinesin II, MKS
  complex) as locations. They now sit on the complex descriptors, and the
  locations are anatomical (axoneme, non-motile cilium, basal body).
• Canonical members often carry only Approved or Uncertain HPA grades. That
  is the main argument for judging atlas calls in a module context.