ADCK5 (Q3MIX3) — review notes

Human ADCK5, 580 aa, chromosome 8, HGNC:21738. Pharos class Tdark.
UniProt name: "Uncharacterized aarF domain-containing protein kinase 5".

Starting hypothesis and how it resolved

The task brief proposed that GO:0004672 protein kinase activity / GO:0006468 protein phosphorylation on ADCK5 would be fold-name-propagated-into-activity errors, on the grounds
that the characterised UbiB members COQ8A/COQ8B turned out not to be canonical protein
kinases.

Outcome: the mechanism is real and demonstrable, but the predicted annotation error is not
in GOA. ADCK5's entire GOA record is four rows and contains no molecular-function term
other than GO:0005515 protein binding. QuickGO returns 0 hits for GO:0004674 with
goUsage=descendants on Q3MIX3. There is nothing to REMOVE or MODIFY on kinase grounds.

The unsupported kinase claim does exist — but one layer up, in UniProt:

• DE RecName: ... EC=2.7.11.- (protein-serine/threonine kinase)
• KW Serine/threonine-protein kinase;
• DR GO; GO:0004674; F:protein serine/threonine kinase activity; IEA:UniProtKB-KW.

…in the same entry that says "The function of this protein is not yet clear." and "It is
not known if it has protein kinase activity and what type of substrate it would
phosphorylate (Ser, Thr or Tyr)." GOA no longer imports keyword-derived (SPKW) annotations,
so the GO record is clean while the UniProt record is not. Reported as a UniProt correction
request rather than a GO action.

Verified rather than assumed, with a control. QuickGO returns 0 human annotations for
GO_REF:0000043 (the Swiss-Prot-keyword pipeline, retired ~April 2026), against 139,714
for GO_REF:0000044 (SubCell) and 1,862 for GO_REF:0000041 (UniPathway) — so the zero is
specific to the keyword pipeline, not an artefact of the query. And GO:0004674 itself is
alive in GOA: PRKACA (P17612) carries 31 annotations to it, by ARBA and by many IDAs. ADCK5
receives it by no route at all.

Family and structural basis (see ADCK5-bioinformatics/)

ADCK5 is a UbiB-family atypical kinase. Stefely et al. defined the UbiB-specific features
[PMID:25498144 "including a unique and invariant KxGQ motif"; "an atypical AAAS motif in an
alanine-rich (A-rich) loop that replaces the canonical glycine-rich (G-rich)
nucleotide-binding loop"], and argued they inhibit protein-kinase function — PMID:25498144.

A MAFFT alignment of the five human UbiB proteins + yeast Coq8p + E. coli UbiB + PKA Cα
(negative control), with every published reference residue asserted before use, shows
ADCK5 retains both inhibitory features and the full catalytic core:

So ADCK5 is neither a canonical protein kinase nor a dead pseudokinase: it has an intact
nucleotide-binding/catalytic core behind a UbiB-type occluded substrate pocket.

The A209 assignment is corroborated two ways — the motif AAAS sits at 207–210 in ADCK5
exactly as it sits at 337–340 in COQ8A, and the COQ8A-anchored and PKA-anchored columns
resolve to the same alignment column (391).

The mirror error, avoided

"UbiB proteins are not protein kinases" would be too strong. COQ8A/Coq8p indeed lack
generic in-trans activity PMID:27499294, but COQ8B has a demonstrated protein substrate: PMID:38425362. The same paper excludes
small-molecule kinase activity for COQ8B PMID:38425362. (Caveat worth carrying: that study used ancestral-sequence-reconstructed COQ8
proteins, so it is a statement about the reconstructed tetrapod enzyme.)

So the right reading for ADCK5 is: generic Ser/Thr kinase activity is unsupported and
architecturally disfavoured; a specific, substrate-restricted activity is open and would
need direct assay. GOA reflects this correctly today by carrying no MF term at all.

Localisation

• UniProt gives only SUBCELLULAR LOCATION: Membrane {ECO:0000305}; Single-pass membrane
  protein {ECO:0000305} — a curator inference from a predicted TRANSMEM 50..67
  (ECO:0000255). That maps to SubCell SL-0162 → the GO:0016020 membrane IEA row.

• The mitochondrion is better supported: MitoCoP high-confidence mitochondrial proteome
  (PMID:34800366, HTP); PMID:25498144; and 17 of ADCK5's 25 distinct IntAct partners from the mitochondrial
  interactome study (PMID:27499296) are UniProt-annotated to the mitochondrion.

• Asymmetry worth reporting, correctly attributed. ADCK1 and ADCK2 receive
  SL-0173 Mitochondrion → GO:0005739 from UniProt; ADCK5 receives only SL-0162 Membrane.
  My first reading was that the difference lay in UniProt's SUBCELLULAR LOCATION line rather
  than in the evidence, since all three carry the same MitoCoP HTP row. That was wrong, and
  checkable in-tree once ADCK1's review merged: ADCK1-uniprot.txt:117 reads
  SUBCELLULAR LOCATION: Mitochondrion {ECO:0000269|PubMed:33988507}, MitoCoP is not cited in
  that entry at all, and ADCK2's entry carries the same ECO:0000269|PubMed:33988507. Both
  paralogs have a dedicated experimental localisation that ADCK5 does not.

• But the reason ADCK5 lacks it is absence of testing, not a negative result — that study
  states plainly PMID:33988507. QuickGO
  returns 0 annotations for Q3MIX3 from that reference, consistent with never having been
  assayed. So the UniProt correction request stands, but on ADCK5's own evidence (MitoCoP HTP;
  17/25 mitochondrial interactome partners; family-wide mitochondrial distribution) rather
  than on a parity that does not hold.

• Submitochondrial assignment (inner membrane?) is not established for ADCK5. Combining
  "mitochondrion (HTP)" with "membrane (predicted TM)" to assert GO:0031966 would be a
  composite claim; left as a suggested experiment instead.

PAINT

PTHR43173 (ADCK5 = SF28, ADCK1 = SF19) has exactly one annotated node,
PTN005148758, seeded by a single yeast protein, MCP2 (SGD:S000004243, itself SF19),
carrying GO:0005743, GO:0007005, GO:0055088. ADCK1 inherits all three; ADCK5 inherits
nothing, because the node sits in the SF19 clade.

ADCK5 is the only human UbiB gene with zero IBA rows — UniProt states it independently:
PAN-GO; Q3MIX3; 0 GO annotations based on evolutionary models. This is the inverse of this
campaign's usual finding: under-reach, not over-propagation. Filed as a question to PAINT,
not as an action, because ADCK5 genuinely sits outside the annotated node's clade and the
node's evidence is a single yeast seed.

GO:0005515 — the NOTCH2NLA rows

Both rows name the same partner, NOTCH2NLA (Q7Z3S9). UniProt records NbExp=4, but expanding
IntAct shows PMID:25416956 logs the interaction three times as three sub-method labels of
one Y2H screen (two hybrid array + two hybrid prey pooling approach + validated two hybrid), and PMID:31515488 adds one more two hybrid array from the same CCSB resource
lineage. MI score 0.67 throughout; no orthogonal assay in any of ADCK5's 54 IntAct records.
Third occurrence of this NbExp trap in the campaign (after ACRV1, ADAMTSL5).

NOTCH2NLA is Secreted/Cytoplasm, human-specific, and functions in neural progenitor
proliferation. Y2H places both proteins in the yeast nucleus and so removes the targeting
constraint.

The compartment argument is an assumption, and is stated as one. It holds if ADCK5's
kinase-like domain faces the matrix, as COQ8A's C-terminus is measured to do PMID:27499294 — but ADCK5's own sidedness has never been measured, and an outer-membrane
anchor presenting the domain to the cytosol is not excluded. That is the same uncertainty that
stops this review proposing GO:0031966, so leaning on the compartment argument while
declining the localisation refinement would be inconsistent. The verdict does not need it: the
method-replication argument stands alone. (Caught by the PR reviewer; conceded.)

Checks that came back negative (recorded so they are not re-run blindly):
- NOTCH2NLA resolves to a reviewed, canonical, full-length Swiss-Prot entry — no
TrEMBL/partial-ORFeome substitution as on ACRV1.
- PMID:34800366 is a proteome-wide localisation census carrying no functional/phenotype
term, so the ACTR8 complex-to-subunit projection failure mode does not apply.
- No retraction, erratum or expression of concern on any cited PMID
(CommentsCorrections/RefType on each cited record, plus Crossref relation/update-to
for the six load-bearing DOIs).

Literature quality

The only ADCK5-specific functional paper is PMID:32277958 (lung cancer, SOX9/PTTG1). Its
abstract is hedged throughout — "showed that ADCK5 might regulate the expression of tumor
oncogene human pituitary tumor transforming gene-1 (PTTG1) by phosphorylating transcription
factor SOX9" — and it reports no in vitro kinase assay; the SOX9 S181 claim rests on
mutagenesis of the substrate plus a motif-match argument, "The serine 181 site of SOX9 is in
a motif that is targeted by ADCK5." Full text unavailable. The affinage record restates this
as fact ("ADCK5 phosphorylates the transcription factor SOX9 at serine 181"); the abstract
does not support that strength. No GOA row rests on this paper, so nothing needed changing —
recorded because the discrepancy is the kind that would otherwise propagate.

affinage gates_passed: True, faith_pct: 50.0, 5 citations, all numeric PMIDs (no
PMID:bio_* preprint ids). Its remaining findings (senescence, JQ1, asthma) are
low-confidence, single-study, and none underpins a GOA row.

Verdicts

No core_functions are asserted. ADCK5's molecular activity is genuinely undetermined —
UniProt says so, Pharos calls it Tdark, and this review found no measurement to replace
that. Per CLAUDE.md the "No core functions defined" warning is left standing rather than
silenced with invented content.

That absence is now recorded positively rather than only as a warning: two knowledge_gaps
entries carry it, a BIOLOGY/WHOLLY_DARK gap for the undetermined activity, substrate and
process, and an ONTOLOGY/MF_DARK gap for GO's inability to express the UbiB
occluded-pocket architecture. Both are grounded in quotes already verified in this review.
(The reviewer's point, and a good one: as prose in suggested_questions the finding is read
once; in knowledge_gaps it feeds the Function Knowledge Gaps project.)

A break-test can be too destructive to discriminate

Worth recording because it is a failure mode this campaign has not named, and it kept a
defect alive across two rounds of this review.

I reported that a vacuity counter had been "moved inside the sentence loop". I had added the
inner increment and left the outer one, so the guard still reported itself exercised whenever
a unit tripped the pre-filter even if no sentence routed — the precise blindness it exists to
detect. The reason it survived my own break-testing is the general point:

the only committed vacuity break-test blanked the topic from the surface entirely, driving
both increments to zero, so it passed identically against the correct and the incorrect
implementation.

Blanking a whole surface proves only that the check reads the surface at all. To certify
"the counter is inside the loop", the mutation has to be that difference: a unit that trips
the pre-filter while no sentence routes. That probe ("COQ8A is a paralog of interest. The row is IDA." — paralog and token present, never in one sentence) reports the guard exercised
under the old implementation and reports vacuity under the new one, so it discriminates.

The mutation must be as fine as the claim. A coarser mutation still goes green, and green
is what makes it feel tested. This sits alongside the other guard failures found here — a
check that failed on perfect agreement, an unreachable branch that read as coverage, a probe
placed where it was easier to catch than the real thing — and it is the subtlest of them,
because the break-test genuinely ran and genuinely passed.

Reconciliation

GOA TSV: 4 data rows. existing_annotations: 4 entries. No collapse; the two GO:0005515
rows share a partner but differ by reference, and the seeder keys on reference, so both
survived.