ACTR8 (ARP8, hArp8) — review notes

UniProt: Q9H981 (ARP8_HUMAN), 624 aa, Swiss-Prot reviewed, PE 1: Evidence at protein level.
Family: [file:human/ACTR8/ACTR8-uniprot.txt "Belongs to the actin family. ARP8 subfamily."]
Structure: [file:human/ACTR8/ACTR8-uniprot.txt "DR PDB; 4FO0; X-ray; 2.60 A; A=34-624."]

GOA snapshot reviewed: 32 TSV data rows → 31 review rows (GOA carries the
GO:0031011/IDA/PMID:21303910 row twice, once from ComplexPortal and once from
UniProt; the seeder de-duplicates it).

1. Provider record: empty, and that is a provider gap, not a literature gap

The affinage record has n_discoveries: 0, citation_count: 0, no findings table and no
citations, and no self_evaluation_pairwise score; the trust gates were clear only vacuously —
there is nothing to gate. Per the campaign
rule, an empty provider record is not evidence that literature is absent, and here it is
demonstrably wrong: the UniProt RN list alone carries 13 references including a 2.6 Å crystal
structure paper with FUNCTION, SUBUNIT, AND ATP-BINDING SITES in its RP line
(PMID:22977180). ARP8 is one of the best-characterised nuclear actin-related proteins.
Nothing in this review rests on the provider record.

Everything below was assembled from UniProt, the GOA TSV, QuickGO, the PANTHER PAINT file
committed in this repo, ComplexPortal, IntAct and 20 primary papers.

Retraction screen: all 20 PMIDs cited here were checked against PubMed publication types and
an explicit retracted publication[pt] OR expression of concern[pt] query over the whole set —
0 hits. No reference in this review is retracted or carries an expression of concern. But
that method is insufficient on its own — see §15, which records one Publisher Correction it
missed and the check that finds them.

2. What ARP8 actually is

Two distinct jobs, and the literature separates them cleanly.

(a) The nucleosome-recognition module of INO80. ARP8 sits with ARP4, nuclear actin and the
HSA domain of the INO80 ATPase in one of the three modules of the human complex
PMID:21303910,
and that module is part of the catalytically competent core
PMID:21303910.
A recombinant human minimal core containing ARP8 slides nucleosomes
PMID:27257055,
PMID:27257055.

The recognition itself is quantified on the purified human protein
PMID:22977180.
Table 2 mixes species and must be read per-species — see §13, where a cross-species join in my
first draft produced an inverted conclusion. The human rows only:

HsArp8 Δ1–33 is closely comparable (nucleosome 62.6 ± 16, nH 2.11 ± 0.4)
PMID:22977180,
PMID:22977180.
The paper reports cooperativity
PMID:22977180,
and this is the biochemical basis of UniProt's dimer statement
[file:human/ACTR8/ACTR8-uniprot.txt "it may act as a nucleosome recognition module within the complex."] —
but note the full-length human value nH = 1.31 ± 0.4 overlaps 1; the unambiguous cooperativity is
in Δ1–33 (2.11 ± 0.4) and the yeast sub-complex (3.13 ± 0.7). The GO:0031491 proposal therefore
rests on the affinity, not on the Hill coefficient.

(b) A mitotic, INO80-independent role on condensed chromosomes.
PMID:18163988,
PMID:18163988,
and decisively
PMID:18163988.
This is why UniProt records
[file:human/ACTR8/ACTR8-uniprot.txt "Note=Specifically localizes to mitotic"] chromosomes,
and why ARP8–ACTR5 co-IP is seen in interphase but not in metaphase-arrested cells.

Consequence for curation: GOA has no term for (b). GO:0007080 mitotic metaphase chromosome alignment is not a descendant of GO:0033044 regulation of chromosome organization
(checked against the QuickGO ancestor list: GO:0007080's ancestors include GO:0051276 but not
GO:0033044), so it is a genuine addition, not a specificity upgrade of an existing row.

3. Residues, both directions — answered from the human crystal structure, not an alignment

The brief's residue check does not need a homology model here: PDB 4FO0 is human ARP8 itself.

ATP pocket: retained, and the activity was measured.
PMID:22977180
ATP is observed, not inferred; UniProt's BINDING features (55, 56, 283..286) carry no
ECO:0000250 similarity tag and come from this structure
[file:human/ACTR8/ACTR8-uniprot.txt "FT BINDING 55"]. The pocket is functional:
PMID:22977180,
PMID:22977180.
The authors bound their own claim, and I keep the bound:
PMID:22977180,
and no ligand they tested stimulated it
PMID:22977180.
Catalytic-activation residues are substituted, not conserved: actin's Q137 → E266^Arp8^,
actin's H73 sensor → R187^Arp8^. So hydrolysis exists but has no demonstrated physiological
trigger. Verdict: ATP binding is established; ATP hydrolysis is measured but functionally
unassigned — I propose the former and file the latter as a question, not a term.

Polymerisation interface: not retained, and structurally explained.
PMID:22977180
This is the positive structural argument that a cytoskeletal-actin term does not belong on
ARP8 — and it is not needed, because no such term is in GOA and PAINT actively blocks it
(section 5).

The pocket is not decorative — it gates DNA binding.
PMID:25299602,
and the effect is lost in pocket mutants
PMID:25299602,
i.e. exactly the residues UniProt annotates as ATP-binding. That is a functional consequence
of nucleotide occupancy, which is more than most ARPs can show.

Sibling consistency (checked, not assumed). ACTR1A and ACTR1B each proposed GO:0043531 ADP binding as NEW; ACTR10 proposed GO:0005524 ATP binding as NEW. Same stated method —
annotate the ligand actually resolved — different observed ligand. ARP8's structure resolves
ATP, so GO:0005524 is the consistent choice, not a divergence.

4. DNA binding: two papers that look contradictory and are not

• 2012, 30-bp duplex, fluorescence titration:
  PMID:22977180
  and the complex's DNA affinity is attributed elsewhere:
  PMID:22977180.
• 2014, long linear/circular DNA and ssDNA, gel shift:
  PMID:25299602,
  PMID:25299602,
  PMID:25299602.

Reconciliation: the 2012 assay used a 30-bp duplex, the 2014 assay long DNA and ssDNA, and
the binding element is the basic N-terminal extension
PMID:25299602
which is dispensable for a 30-mer but required for long duplex
PMID:25299602.
Independent yeast corroboration of the same element (labelled as yeast, not asserted for human):
PMID:30120252

So the well-founded MF is the specific one, GO:0003697 single-stranded DNA binding, not a
bare GO:0003677 DNA binding: ssDNA is the preferred ligand in a direct competition assay
(dsDNA vs 3′-overhang vs ssDNA), and it is the ligand whose relevance to DSB-end resection the
authors argue. Neither term is in GOA.

5. PANTHER node PTN000234048: the nuclear-ARP/cytoplasmic-actin question, both directions

This is the check that damaged ACTL8 (a divergent actin mis-placed inside the β/γ-actin
subfamily) and ACTR10 (a nuclear term transferred in from another subfamily). For ACTR8 both
directions come back clean, and the second direction comes back better than clean.

Census, via QuickGO withFrom=PANTHER:PTN000234048 (paged; numberOfHits = 348, all 348
fetched): 87 distinct gene products, each carrying exactly the same 4 terms (GO:0003729,
GO:0006302, GO:0006355, GO:0031011). Every named member is an ARP8 orthologue —
ACTR8/Actr8/actr8/arp8/ARP8/arp-8/arpG, plant ARP9 — across plants, fungi,
metazoa, Giardia and Paramecium. Zero conventional actins: a regex sweep for
ACTB|ACTG*|ACTA*|ACT1|ACT2|POTE|ACTL*|ACTR[0-7,9] over the member symbol list returns an
empty set. So this node is an ARP8-specific orthogroup, not the broad actin node, and no
cytoplasmic-actin cell biology can enter ACTR8 through it.

The reciprocal direction is the interesting half. The repo's own PAINT file shows PAINT has
placed an explicit negative at this node:
[file:interpro/panther/PTHR11937/PTHR11937-paint.tsv "PTHR11937 PTN000234048 GO:0005200 F IRD true PANTHER:PTN000940351"]
— IRD, NOT-flagged, against PTN000940351, which is the conventional-actin node
[file:interpro/panther/PTHR11937/PTHR11937-paint.tsv "PTHR11937 PTN000940351 GO:0005200 F IBD false SGD:S000001855|UniProtKB:P61160"]
(donors include yeast ACT1 and human ACTB P60709; 969 annotations over ≥200 gene products, all
GO:0005200). PTHR11937 carries 9 GO:0005200 rows in total, of which 8 are IRD
negatives at divergent nodes and 1 is the IBD positive at the actin node.

So GO:0005200 structural constituent of cytoskeleton is deliberately blocked from reaching
ARP8, and that block agrees exactly with the human crystal structure's "inability to polymerize"
finding. ACTL7A and ACTL7B each had to REMOVE a GO:0005200 TAS row; ARP8 never acquired one.
This is PAINT working correctly on the same family where it failed for ACTL8 — worth reporting
upstream as the positive control for that fix.

The reciprocal direction: is ARP8 a donor in the nucleus leak that hit ACTR10?

No. ACTR10 merged into main mid-review (#2274) with its GO:0005634 nucleus IBA set to
REMOVE as a paralog transfer, so it was worth asking whether ARP8 — a genuinely nuclear ARP —
was the source. Reading the merged review's source_entities, ACTR10's nucleus donors are
SGD:S000004636 (S. cerevisiae ARP9, a SWI/SNF and RSC subunit), CGD:CAL0000196900
(Candida ARP9), MGI:MGI:1343051 (mouse Actl7a) and UniProtKB:Q57ZL0 (an unreviewed
Trypanosoma entry), at node PTN008986520 — a different and deeper node than
PTN000234048, with no ARP8 orthologue among them. Confirmed against ACTR10's own GOA
WITH/FROM field, which lists exactly those five tokens.

So the nuclear leak in PTHR11937 runs through the ARP9/ACTL7 side of the family, not the
ARP8 side, and ARP8's own nuclear terms are independently earned. Both halves of the brief's
question answered, and both negatives recorded.

Sibling-review consistency, checked row by row

None of these is an inconsistency, and each difference is traceable to the evidence rather than
to the rule:

• Protein binding — the shared rule is decide per partner. ACTR8's two partners (ACTR5,
  UCHL5) are both co-members of its own ComplexPortal complex, each established by a targeted
  experiment; ACTR10's and ACTL7B's were screen singletons. Same rule, different partners.
• Nucleotide ligand — the shared rule is annotate the ligand actually observed. ARP8's own
  structure resolves ATP, so IDA; ACTR10 had to reconstruct its site by transferring β-actin's
  contacts across an alignment, so ISS; the Arp1 paralogs resolve ADP, so GO:0043531. This is
  the "convergent method, divergent term" pattern — no reconciliation needed.
• GO:0005200 — ACTR10 keeps it because Arp11 really is a dynactin filament constituent;
  ACTL7A/7B lose it because their filament interfaces are not conserved; ARP8 never had it
  because PAINT blocked it, in agreement with the human structure. Three outcomes, one criterion.

Per-node term assignment for PTN000234048 (all 5 rows, from the committed PAINT file):

Two of the four positives are self-referential (human Q9H981 is itself in the WITH/FROM
list) → NO_FAILURE_CORE, a PAINT curator judging the function core for this gene.

6. Donor evidence, queried per term (not merely resolved)

All WITH/FROM accessions resolved with size=5, primaryAccession checked against the request,
and Swiss-Prot/TrEMBL status printed. Every one is a reviewed entry; no dead accessions, no
TrEMBL name-provenance traps.

No paralogs in the donor set. All are true ARP8 orthologues. That is unusual for this
campaign and it means every IBA row here has ortholog-strength support available in principle —
the questions are only which term and how many donors carry it.

Donor evidence for the term each donated, via QuickGO per-accession + per-GO queries:

• GO:0031011 — yeast ARP8 IPI ×3 (PMID:10952318, PMID:24034245), Drosophila Arp8 IDA+IPI
  (PMID:16618800), S. pombe arp8 IDA (PMID:19040720) + IPI (PMID:19933844), human
  IDA ×3. Four independent experimental donors plus the target's own IDA. Unambiguous.
• GO:0006355 — Drosophila Arp8 IMP (PMID:16618800); human self.
• GO:0006302 — yeast ARP8 IMP (PMID:23644470).
• GO:0003729 — yeast ARP8 IDA (PMID:20844764) and nothing else, in any species.

7. The two findings worth reporting upstream

7a. GO:0003729 mRNA binding — one yeast RIP-chip hit, propagated to 87 eukaryotes

I expected this to be an unvalidated screen artefact and it is not — checking rather than
assuming changed the verdict. Yeast ARP8 is in the paper's confirmed set:
PMID:20844764
PMID:20844764
231 targets at FDR ≤ 0.01 % — the Methods define the cut-off as a SAM-calculated false
discovery rate "less than or equal to 0.01%", i.e. 100-fold stricter than a bare 0.01 — and only
PMID:20844764
So the SGD IDA is sound and SOURCE_EVIDENCE_WEAK would be factually wrong.

What is wrong is the propagation, on three independent grounds:

1. One donor, one study, whole eukaryotic clade. GO:0003729 is IBD from SGD:S000005667
   alone, added 2025-08-05, and now sits on all 87 members, no member outside S. cerevisiae
   having ever been assayed for RNA binding. A single donor is not by itself the objection —
   claiming otherwise would contradict this review's own GO:0006302 row. Donor counts at the
   node are GO:0031011 4, GO:0006355 2, and GO:0003729 and GO:0006302 one each, yet
   GO:0006302 is accepted. What separates them is target-side evidence: human ARP8 has direct
   DSB-repair data and no RNA data at all.
2. The assay cannot distinguish RNA from DNA binding. The primary screen is
   PMID:20844764
   — RNA in, RNA out. No DNA competitor was included anywhere in the design. For a protein whose
   nucleic-acid-binding element is a basic, disordered N-terminal extension and whose
   demonstrated preference is for single-stranded nucleic acid, "binds polyanion" and "binds
   mRNA" are not separable by this experiment.
3. The binding element differs in size between donor and target — but read the whole
   sentence. My first draft of this point overstated it, and the correction is worth recording
   because it is the exact failure mode the campaign brief warns about. The tempting quote is
   "It is rather short in human Arp8 compared to yeast Arp8". Continuing to the end of the
   qualifying clause:
   PMID:22977180.
   So the human protein retains a substantial basic extension. Yeast ARP8 is 881 aa against
   human 624 aa with much of the difference in this region, which makes the transfer less safe —
   it does not make it impossible. Grounds (1) and (2) carry the argument; (3) only supports
   it. A verbatim quote truncated one clause early would have read as a much stronger finding
   than the data support.

Verdict MARK_AS_OVER_ANNOTATED, not REMOVE: the donor annotation is real and I have no human
negative — no one has assayed human ARP8 with RNA. REMOVE would require a positive argument I
cannot supply, and an absence of human RNA data is an absence, not a finding.
root_cause: PROPAGATION_BAD (source sound, transfer unsafe), which is what my own reason
argues. SOURCE_EVIDENCE_WEAK is deliberately absent from failure_modes because my own
analysis contradicts it.

GRANULARITY_MISMATCH — which half of the definition. The enum definition has two arms:
"Parent term is true but uninformative, or child term overstates specificity." The brief's
caveat (donors must agree before invoking it) applies to the first arm. I invoke the second
arm here: the donor's own term names a ligand class its assay did not discriminate. Each row
using GRANULARITY_MISMATCH now states which arm it means — rows GO:0033044 and GO:0051052
use the first (redundant true parent), rows GO:0003729 and GO:1904507 the second.
FUNCTIONAL_DIVERGENCE is deliberately not claimed on the mRNA-binding row: a change of
ligand class in the human protein has not been demonstrated, only that its characterised
ligand is a different one, and asserting divergence I have not measured would be the same error
as accepting an activity from a fold name.

7b. A complex-level phenotype became a subunit "IMP" and was then transferred across species

Five of ACTR8's 31 rows (GO:0000723, GO:0006282, GO:0045739, GO:0045995, GO:1904507)
are IEA/GO_REF:0000107 from mouse Actr8. Tracing the mouse side:

• All five mouse Actr8 annotations come from one paper, PMID:23979016, and all five are
  assignedBy: ComplexPortal.
• Querying reference=PMID:23979016 in QuickGO returns 80 annotations = 16 entities × 5
  terms: ComplexPortal:CPX-878 (mouse INO80 complex) plus all 15 subunits, byte-identical
  term/evidence sets on Ino80, Ino80b/c/d/e, Ruvbl1/2, Yy1, Tfpt, Nfrkb, Mcrs1,
  Uchl5, Actl6a, Actr5 and Actr8.
• The paper's experiment is an Ino80 knockout:
  PMID:23979016
  The embryonic phenotype behind GO:0045995 is an Ino80-null phenotype:
  PMID:23979016
  and the telomere conclusions behind GO:0000723/GO:1904507 are stated for the complex:
  PMID:23979016
  Actr8 is not named in the abstract, and no Actr8 mouse knockout exists.

ComplexPortal's own layer is a documented convention — it annotates the complex and projects to
members, and the complex context is recoverable from CPX-878. The defect is the second
step: Ensembl Compara consumes the projected subunit annotation as if it were gene-specific
mouse experimental evidence and re-projects it onto human ACTR8, at which point the complex
context is gone. Two inferences stacked, presented as one IEA from an ortholog. Hence
root_cause: EVIDENCE_CIRCULAR_OR_REDUNDANT with CIRCULAR_PROPAGATION + ROLE_CONFLATION:
the transfer depends on another inference, not on independent mouse-gene evidence.

Split by whether the term survives on ACTR8's own evidence:

• GO:0006282 / GO:0045739 (regulation / positive regulation of DNA repair) — do survive
  independently, on human ARP8-specific data
  PMID:20971067
  and PMID:25299602.
  → KEEP_AS_NON_CORE, with the bad route recorded rather than the term deleted.
• GO:0000723, GO:1904507, GO:0045995 — have no ARP8-specific support in any organism.
  → MARK_AS_OVER_ANNOTATED. Not REMOVE: ARP8 is a core INO80 subunit and I cannot show it is
  dispensable for these complex functions. The honest statement is that the specificity is
  unearned, not that the biology is false.

Generalisation for PAINT/GOA: wherever ComplexPortal projects a complex phenotype to N
subunits, Ensembl Compara multiplies it by the number of species with 1:1 orthologues. For
CPX-878 alone that is 15 subunits × 5 terms; ACTR8 is one of 15 human genes receiving the same
five rows by the same route. Stated once here, for all affected genes.

8. Interaction rows: the ACRV1/ACRBP checks, run and reported including the nulls

IntAct findInteractions/Q9H981: totalElements 82. Methods: anti tag coip 55,
anti bait coip 18, proximity-dependent biotin identification 4, tap 2,
2h fragment pooling 1, pull down 1, affinity chrom 1.

• Not one screen counted several ways. The 82 records span 14 distinct publications, and
  the two GOA partners are supported by targeted experiments, not only by maps: ACTR5 by
  anti bait coip in PMID:18163988 at MI-score 0.67 (physical association), UCHL5 by the
  dedicated PMID:18922472 study. So the ACRV1 pattern (one Y2H logged under three sub-method
  names) does not apply here.
• Partner accessions are canonical. Q9H9F9 = reviewed ACTR5, 607 aa = canonical length;
  Q9Y5K5 = reviewed UCHL5, 329 aa = canonical length. No TrEMBL clones, no ORFeome fragments,
  no dead accessions. This check came back negative and is recorded as such.
• Both partners are bona fide complex co-members. ComplexPortal CPX-846 (human INO80
  chromatin remodeling complex) has 15 participants and both Q9H9F9 and Q9Y5K5 are among
  them, alongside Q9H981. So every GO:0005515 row on ACTR8 is a within-complex contact whose
  informative content is already carried by GO:0031011.
• A result worth recording, and a correction to my own first draft. I initially wrote that
  INO80E is "the most reproducibly detected ACTR8 partner in IntAct". That was false — an
  unverified superlative built on true constituents, which is the campaign's dominant failure
  shape. Ranking partners by distinct supporting publications:

YY1 ties INO80E on publication count and beats it on MI-score, so the superlative was wrong.
The corrected statement is a stronger finding: both best-attested partners are absent
from GOA, and the two that are recorded rank 6th and 3rd. The omission of YY1 is the sharpest
case — PMID:18026119 is already cited on this gene for the Ino80-complex IDA, and YY1 is the
bait that paper used, so the contact rests on a reference GOA has in hand. Not a defect in the
strong sense (curators select what to annotate), but it settles the direction: these six rows
under-represent the interaction data.

Handling: all six KEEP_AS_NON_CORE, matching how the merged siblings treated real, replicated
partners (ACTR1A, ACTR1B, ACTL7A all used KEEP_AS_NON_CORE; MARK_AS_OVER_ANNOTATED was
reserved for singleton screen hits in ACTL8/ACTL7B). No divergence from sibling practice.

9. Where GOA's molecular-function coverage stands

QuickGO for Q9H981, aspect=molecular_function: 7 rows total — 6 × GO:0005515 protein binding and 1 × GO:0003729 mRNA binding (IBA). An explicit query for GO:0005524 with
goUsage=descendants returns total: 0.

So the entire molecular-function record of a protein with a 2.6 Å ATP-bound crystal structure, a
measured ATPase, quantified nucleosome and histone-tetramer affinities and a demonstrated ssDNA
preference consists of "protein binding" six times plus one cross-kingdom RNA inference. The
GO:0005524; F:ATP binding; IEA:UniProtKB-KW line still present in the UniProt DR block is
not in GOA — consistent with the withdrawal of Swiss-Prot-keyword-derived annotations
(GO_REF:0000043) from GOA; GO:0051301 cell division, GO:0006310 DNA recombination and
GO:0006351 DNA-templated transcription disappeared from ACTR8 the same way.

That makes the MF gap the substantive deliverable, and it is filled from human experiments on
the human protein, not from the fold name. Proposed as NEW:

GO:0000793 is additive, not a replacement for the existing GO:0005694 chromosome row:
ARP8 is also on interphase chromatin as an INO80 subunit (it is required for INO80 recruitment to
laser-induced damage in interphase), so the general term remains the correct cover and the
condensed-chromosome pool is an extra.

GO:0016887 ATP hydrolysis activity is deliberately not proposed. It was measured on the
human protein, but the authors themselves place it at "extremely weak … in a monomeric state"
with no ligand able to stimulate it, so its physiological status is unresolved; it belongs in
suggested_questions, not in the annotation set. Accepting a hydrolase term off a measured but
functionally unassigned trace activity is the mirror of accepting one off a fold name.

10. Complex-subunit schema decisions

• INO80 membership goes in core_functions[].in_complex (GO:0031011), never locations —
  the latter is bound to a location enum and would fail validation. Verified GO:0031011 is
  present in cache/enums/goproteincontainingcomplexenum_*.csv.
• molecular_function vs contributes_to_molecular_function: ARP8's binding activities were
  measured on the isolated purified protein (Kd values for nucleosome, (H3–H4)₂, H2A–H2B,
  DNA; ssDNA gel shifts on recombinant ARP8 alone), so ARP8 independently enables them and
  molecular_function is correct per the schema's enables definition. The one activity it does
  not independently enable is nucleosome remodelling — that is the INO80 ATPase's — so the
  remodelling core function is stated as a BP (GO:0006338) with the complex in in_complex,
  and no remodeller MF is claimed for ARP8.

11. Rows where I deliberately did not upgrade specificity

• GO:0006355 regulation of DNA-templated transcription (IBA) stays as-is even though ACTR8 has
  its own IMP to the child GO:0045893. The complex works in both directions — it
  activates PMID:27641337
  and it represses PMID:26340092.
  A direction-neutral parent is the correct LCA, so GRANULARITY_MISMATCH does not apply: the
  donors and the target's own data do not agree on sign.
• GO:0005694 chromosome (IEA) → not modified to GO:0000793; see §9.
• GO:0051052 regulation of DNA metabolic process (ARBA) is true but wholly subsumed by the
  separately annotated GO:0006275 and GO:0006282. Kept non-core rather than flagged as
  over-annotation: a redundant true parent is not an over-annotation.

12. Three claims of my own that a self-audit retracted

Recorded because the corrections are the useful part, and because each is a composite claim —
every constituent quote verbatim, the join unsupported. None would have been caught by quote
validation.

1. "the most reproducibly detected ACTR8 partner in IntAct is INO80E" — false; YY1 ties on
   publication count and beats it on MI-score. See §8; the corrected version is a stronger
   finding.
2. "ARP8 was the only INO80 subunit whose depletion abolished recruitment" — the source says
   PMID:25299602.
   The qualifier is "tested". Dropping it converts a screen result into a claim about every
   subunit. Restored in all three places the claim appears (the GO:0006302 IBA row, the
   GO:0006282 row, and core function 2), and the sentence is now quoted rather than paraphrased.
3. "the only reported function of a nuclear actin-related protein detached from its remodelling
   complex" — an unverifiable negative over the whole nuclear-ARP literature. Softened to
   "rarely reported … one of the few instances with genetic controls behind the claim", with the
   limitation stated inline.

The audit that found these was a grep for superlatives and absolutes (the most, the only,
never, none of, no one has, all N) over every string in the review plus these notes,
followed by re-deriving each hit from its source. Twelve numeric claims were re-derived
independently against QuickGO, ComplexPortal, IntAct and the committed PAINT file — node census
348/87/4 terms, MF row count 7, GO:0005524 count 0, PMID:23979016 80 = 16 × 5, CPX-846 15
participants, IntAct 82 records / 14 publications, PTHR11937 9 GO:0005200 rows = 8 IRD + 1 IBD,
and the two ontology-ancestry claims — all confirmed. The three failures above were all
qualitative superlatives, not numbers.

13. The error the reviewer caught: a cross-species join inside one table

Worth recording in full, because it is the single sharpest instance in this review of the failure
class the campaign says dominates — every constituent number verbatim and correct, the join
invalid — and no mechanical check could catch it.

I wrote, in the GO:0031491 NEW reason:

"the isolated Arp8-Arp4-actin-HSA sub-complex binds nucleosomes at 64 nM, so ARP8 accounts for
essentially all of the module's nucleosome affinity."

Both numbers are real. The inference is not, because PMID:22977180 Table 2 is a mixed-species
table and the paper says so in its own methods sentence: it assayed "full-length and
N-terminally truncated human Arp8 as well as yeast Arp8, Arp4 and the Arp8–Arp4–actin-HSA
sub-complex I". The relevant rows, 207-bp nucleosome:

I had set the human 51 nM beside the yeast module's 63.6 nM. Matched within species the
comparison inverts: the module binds ~4.9× tighter than ScArp8 alone, so far from ARP8
accounting for "essentially all" of the module's affinity, assembly increases it. ARP8's
fractional contribution simply cannot be derived from these data, and the defensible statement is
the authors' own hedged one, which I already quote.

Three further species labels were missing for the same reason, all now fixed and each verified
against the source:

Generalisable lesson: when a table's header enumerates species (Hs/Sc), treat every
cross-row comparison as cross-species until proved otherwise, and re-derive any ratio within one
species. The affected verdicts did not change — GO:0031491 IDA with enables rests on
HsArp8 at 51 ± 9.6 nM measured on the purified human protein — which is exactly why the error
survived my own audit: it sat in the justification, not in the term, the evidence code or the
qualifier, and my superlative sweep (§12) was looking for absolutes rather than for species.

Two smaller reviewer corrections accepted on the same pass:

• FDR units. PMID:20844764's cut-off is FDR ≤ 0.01 %, not 0.01 — its Methods define
  "a SAM-calculated False Discovery Rate (FDR) less than or equal to 0.01%". This makes the donor
  annotation stricter than I had described, which strengthens the finding that the source is
  sound and the fault lies in the propagation.
• Boilerplate. The shared protein-binding and Compara blocks were ~200 and ~230 words repeated
  across 6 and 5 rows respectively, burying the per-row reasoning. Both are now compact pointers
  into §8 and §7b, with each row keeping only what is specific to it (the 17 records from
  PMID:16230350, the MI 0.67 targeted co-IP, NbExp=5, the YY1/INO80E ranking, the negative
  regulatory arm, the Ino80-null embryonic phenotype, the three conjoined claims).

14. Two parallel PRs, one new term, one duplicate — a gate that cannot fire per-PR

Found while re-running the gates for this follow-up. cache/go/terms.csv on main carried
GO:0031011 twice:

ARP5 and ARP8 are both INO80 subunits, so both reviews needed the same new term for
core_functions.in_complex. Each branch's duplicate gate
(cut -d, -f1 cache/go/terms.csv | sort | uniq -d) passed in isolation, because within either
branch the term appeared once — and they landed in different positions, so git auto-merged both
without a conflict. The duplicate exists only in the merge result.

This is a new variant of the known duplicate failure mode: the documented one is self-inflicted
within a single branch (hand-insert, then a later just validate re-appends because the appender is
blind to the inserted row). This one is inter-branch, and no per-PR check can see it — the
per-branch gate is not wrong, it is simply looking at the wrong artefact. The EOF duplicate is
removed here, keeping the sorted-position row, so the term appears exactly once.

Two generalisations worth carrying:
1. Related genes reviewed in parallel are the high-risk case, because they are precisely the
ones that need the same new terms. Whenever a sibling PR is in flight, expect a terms.csv
collision on whatever complex or activity term they share.
2. The duplicate check belongs in CI on main, not only in each PR, since only the merge
result can exhibit the fault. A per-PR approximation is to re-run the check after
git merge origin/main immediately before pushing — which this review did on three merges, but
ACTR5 had not yet merged at that point, so the collision was still invisible. Filed as
issue #2294 rather than fixed here,
because the workflow is shared CI config outside gene-review scope; the issue records that
main already carries two pre-existing duplicates (GO:0001675, GO:0009566) which a bare
"must be empty" assertion would trip over, and that the file must not be re-sorted.

15. The correction screen used the wrong instrument

Prompted by a cross-gene note that provider records cite corrected papers unflagged. Re-running
the screen properly found one item my round-1 method could not have found, and the method is
the transferable part.

What round 1 did: an esearch over the 20 cited PMIDs AND (published erratum[pt] OR corrected and republished article[pt] OR retracted publication[pt] OR expression of concern[pt]).
Re-run now, that query returns Count: 0.

Why it is wrong: those publication types sit on the erratum record, not on the article being
corrected. Querying the cited set therefore asks "is any paper I cite itself an erratum?", which
is never the question. The right field is CommentsCorrections/RefType on each cited article,
read from efetch&retmode=xml, matching ErratumIn, RetractionIn, ExpressionOfConcernIn,
CorrectedAndRepublishedIn, RepublishedIn, PartialRetractionIn.

What that found, across all 20:

PMID:41039152 is a Publisher Correction (Nature 646:E16, Oct 2025) to "Multimodal cell maps
as a foundation for structural and functional genomics" — a publisher-side correction, not a
data- or figure-integrity notice, and the article is not retracted. The single claim this review
rests on it is an ACTR8–UCHL5 association independently established by PMID:18922472 and by
CPX-846 membership, so nothing changes. is_invalid is deliberately not set: that flag is
for retracted or replaced references, and using it here would misrepresent a typesetting-class
correction as an integrity problem.

Rule to carry forward: screen corrections by reading CommentsCorrections on each cited
article, never by a publication-type query over the cited set. The two CommentIn hits also
show why the RefType must be matched rather than merely counted — a commentary is not a
correction.

The correction record itself is now cached at publications/PMID_41039152.md, so its existence and
its Erratum for linkage to PMID:40205054 are verifiable from the repository rather than resting
on a network call — a reviewer flagged that it was otherwise uncheckable offline, which was fair.
It is deliberately not added to the references: list: it is a correction notice, not a source
of any claim here, and listing it would imply this review draws evidence from it. The consequence,
also deliberate, is that PMID:41039152 appears only in review_notes prose and so will not be
picked up by a programmatic sweep over references — the right place to catch it is the
CommentsCorrections scan described above, run against each cited article.

16. Two cross-gene claims checked, and one refuted

Both arrived as cross-gene guidance while this review was in flight. Recorded with outcomes
because a check whose result is not written down reads the same as a check never run.

Confirmed, and it distinguishes ACTR8 from ACTR5. PMID:25016522's ComplexPortal-assigned
rows (GO:0006275, GO:0060382) are ACCEPTed here rather than kept non-core, on the grounds
that the paper assayed ARP8 itself. Counting occurrences in the cached full text settles it:
27 × "Arp8", 0 × "Arp5"/"ACTR5". The paper has a dedicated results section "Effect of Arp8
depletion on recovery after replication stress", an esiRNA against human Arp8 (485–916,
NM_022899.3), its own qPCR primers, a 4-fold increase in discontinued forks by fibre labelling,
and the conclusion that "depletion of the Arp8 subunit had the same consequences as Ino80
deficiency". So the same rows that are complex-level projections for ARP5 are gene-level evidence
for ARP8 — the asymmetry is real and is why the two reviews should not be harmonised on these
rows.

Refuted. A suggestion that ARP8 should take no nucleotide term without a structural check,
because it has "1 of 5 catalytic positions with no resolved nucleotide". The structural check was
done, and it is stronger than an alignment: human ARP8 has exactly one PDB entry, 4FO0,
whose own deposited title is "Human actin-related protein Arp8 in its ATP-bound state" and whose
ligand list contains ATP and MG (the physiological Mg²⁺ counter-ion). Add the omit map
computed for the bound ATP, an ATPase measured on the human protein, and ATP occupancy gating DNA
binding through the annotated pocket residues, and GO:0005524 is IDA-grade. The residue
observation is nonetheless correct and is already reflected: actin's catalytic Gln137 and sensor
His73 are substituted in ARP8 (Glu266, Arg187), which is precisely why GO:0016887 ATP
hydrolysis activity is withheld. Binding and hydrolysis are separate claims and the review
separates them; "no nucleotide term at all" does not follow.

Null result, recorded as one. A concern that the seeded stub can silently collapse duplicate
GO:0005515 rows, so annotations must be counted against the GOA TSV rather than the stub. Checked
per (term, evidence, reference) key: all 31 distinct GOA keys are present in the review, all
6 GO:0005515 rows survive with their own verdicts, and no review row lacks a GOA counterpart.
The only multiplicity difference is the GO:0031011/IDA/PMID:21303910 row that GOA banks twice
(once ComplexPortal, once UniProt), documented at the top of these notes. No collapse occurred
here — but the check is what establishes that, not the absence of a symptom.

17. Core-vs-non-core rule applied consistently

ACCEPT where the supporting experiment manipulated or measured ARP8/Actr8 itself;
KEEP_AS_NON_CORE where the only evidence is a complex-level manipulation (INO80 knockdown,
CPX-846/CPX-878 projection) or a redundant general parent. Applied uniformly, this puts
GO:0006275 and GO:0060382 in the ACCEPT column (Arp8 was assayed:
PMID:25016522)
and GO:0033044, GO:0045893, GO:0051726 in the KEEP_AS_NON_CORE column (INO80 RNAi only).
The rule is stated here so the pairing of a core parent (GO:0006355) with a non-core child
(GO:0045893) reads as deliberate rather than inconsistent.