ACTR10 (Q9NZ32) — review notes

Human ACTR10 / ARP11 (also ACTR11, ARP10, hARP11), 417 aa, Swiss-Prot since 2003,
PE 1: Evidence at protein level. Reviewed as part of the PAINT + affinage campaign.

1. What the protein is

ACTR10 is the pointed-end capping subunit of dynactin, the ~1.2 MDa activator of
cytoplasmic dynein-1. UniProt states what that end does
[file:human/ACTR10/ACTR10-uniprot.txt "important for binding dynein-dynactin cargo adapters."]
and lists ACTR10 in the four-subunit pointed-end module
[file:human/ACTR10/ACTR10-uniprot.txt "subunits: ACTR10, DCNT4, DCTN5 and DCTN6."],
with location Cytoplasm, cytoskeleton and family assignment
[file:human/ACTR10/ACTR10-uniprot.txt "SIMILARITY: Belongs to the actin family."].
Both the FUNCTION and SUBUNIT lines are ECO:0000250|UniProtKB:I3LHK5, i.e.
by similarity to pig ACTR10 — the accession behind the pig dynactin cryo-EM series.

The protein was discovered as the fourth, previously unrecognised subunit of dynactin's
pointed-end complex, in bovine brain dynactin
PMID:10525537.

2. The scientific crux: fold retained, one interface lost

ACTR10 is actin-related — Pfam PF00022 (Actin), CDD cd10207 ASKHA_NBD_Arp10,
PANTHER PTHR11937 (ACTIN). The campaign's most repeated defect is a fold or domain
NAME propagating into GO as an activity, and the mirror-image error is asserting
"fold without function" without checking residues. Both were tested
(ACTR10-bioinformatics/RESULTS.md, sections A–C). The answer is not one-sided:

The nucleotide site is intact and occupied. Eckley et al. predicted this from
sequence in 1999
PMID:10525537,
and the structures bear it out. A nucleotide is modelled in the Arp11 chain in
19 of 24 deposited PDB entries that contain ACTR10, in the same maps where the
Arp1 and β-actin chains of the same filament carry ADP — the strongest single
statement being the methods line of the 3.37 Å pointed-end structure
PMID:36071160
β-actin and the Arp1 subunits. Computing contacts from the 7Z8M coordinates gives
18 Arp11 residues within 4 Å of that ATP, laid out canonically for the actin fold
(G21-E22-A23-F24…K26 ≡ actin G13-S14-G15-M16…K18; G142 ≡ actin G156; K211 ≡ actin
K213; G307-G308-T309 ≡ actin G301-G302-T303). 17/18 are identical in human ACTR10,
and 0/17 of β-actin's own ADP contacts fall in an alignment gap. So the site has
not decayed.

The polymerisation interface has gone. Eckley et al.'s specific prediction was a
large deletion of actin residues 38–57, the subdomain-2 loop on the pointed-end face
PMID:10525537,
against a conserved barbed-end face
PMID:10525537,
whence the conclusion
PMID:10525537.
In the alignment, 20/20 of actin positions 38–57 align to a gap in human ACTR10 —
the prediction is confirmed exactly, 26 years on. The cryo-EM structures give the
mechanism
PMID:25814576
and
PMID:25814576,
with Arp11 the only capping element at that end
PMID:25814576.
Biochemically Arp11 binds Arp1 directly
PMID:12857853,
and in cells has no free pool
PMID:12857853.
That second sentence must not be read as "Arp11 cannot bind actin" — the immediately preceding
clause reports the opposite result
PMID:12857853,
so the statement is about the absence of a free Arp11 pool in cells, not about binding capacity.
The same paper raises an Arp2/3-like nucleation role as a hypothesis while arguing Arp1 needs no
nucleator
PMID:12857853,
which is why no nucleation or polymerisation term is annotated here: not demonstrated, rather
than excluded. See §14 — and note that the clause refuting the stronger claim was in the cached
abstract all along, two lines above the sentence I quoted.

Nucleotide binding is a defining property of the actin fold, not of actin alone
PMID:42439233,
so retention of the site in a divergent member is not itself surprising — what is
notable is that no one has annotated it.

Curation consequence. GOA has no nucleotide-binding annotation for ACTR10 —
so the fold-derived over-annotation the campaign usually finds is absent here, and
what is present instead is the opposite defect: a real, structurally observed
nucleotide site with no annotation at all. Correspondingly, no polymerisation or
actin-nucleation term is annotated, which is right. The honest caveat is that the
nucleotide is modelled into cryo-EM density in pig Arp11; the two human dynactin
structures (9B7J, 9B85, 3.47–3.49 Å) model ADP on the Arp1 chains and AMP-PNP on
β-actin but leave chain J empty. A human annotation therefore belongs as ISS from
I3LHK5, not IDA, and no hydrolysis or exchange has ever been assayed.

3. Function: capping, complex integrity, cargo coupling

Three strands, of quite different strength.

Structural (strongest). Arp11 caps the pointed end and is the platform on which the
rest of the pointed-end module assembles
PMID:25814576,
PMID:33734450.
Arp11 is recruited by a composite site made by the terminal Arp1 and β-actin subunits
PMID:25814576,
and is the most divergent family member
PMID:25814576.

Complex integrity (functional, monkey Cos7 RNAi). Yeh & Schroer showed Arp11 is
required for dynactin to exist at all
PMID:22918948,
which is why Arp11 knockdown is used as a proxy for whole-dynactin loss
PMID:22918948;
cross-linking places Arp11 at the centre of the module
PMID:22918948,
PMID:22918948.

Cargo coupling (zebrafish genetics). A forward screen gave actr10 mutants with a
cargo-selective retrograde defect
PMID:28414272,
traced to the linkage step
PMID:28414272,
with a separation-of-function construct
PMID:28414272.
Consequence for organelle homeostasis
PMID:33376159.
A second actr10 allele shows the generic dynactin phenotypes too
PMID:29073112,
so the cargo-selectivity of the nl15 allele should not be over-read as ACTR10 being
a mitochondria-only subunit.

Note the tension worth flagging to experts: Yeh & Schroer report Arp11 loss destroys
dynactin, whereas the zebrafish mutant has a cargo-selective defect. Either the
zebrafish allele is hypomorphic for the structural role, or vertebrate dynactin
tolerates partial Arp11 loss better than Cos7 dynactin.

4. Human-specific evidence, and a missing IDA

Human dynactin was purified and solved for the first time in 2025 using a Chlamydia
effector as an affinity handle
PMID:40186871,
after AP-MS from HEK293T recovered
PMID:40186871.
Human ACTR10 (Q9NZ32) is chain J of the deposited models 9B7J and 9B85. GOA still
carries GO:0005869 dynactin complex for human ACTR10 only as IBA; the annotation
is correct but under-evidenced, and an IDA from PMID:40186871 is now available.
Mouse Actr10 already has the IDA (from PMID:10525537), so human lags its own ortholog.

5. PAINT source resolution — the decisive check

Full table in ACTR10-bioinformatics/RESULTS.md section A. Summary of what each IBA
row was propagated from.

term sources verdict
GO:0005869 dynactin complex mouse Actr10 (Q9QZB7, IDA), pig ACTR10 (I3LHK5), C. elegans arp-11 (Q9GYR2, TrEMBL, IDA) true orthologs → sound
GO:0098958 retrograde axonal transport of mitochondrion zebrafish actr10 (Q7ZVU0, TrEMBL, IMP) true ortholog, real loss-of-function experiment → sound
GO:0005200 structural constituent of cytoskeleton 10 protein sources, 10/10 Swiss-Prot with own experimental evidence: mouse/rat ACTG1, yeast ACT1, 2 Dictyostelium actins, human ACTR2 + ACTR3, yeast ARP1 (centractin), yeast ARP10 mixed: mostly polymerising actins, but includes the dynactin filament subunit and the true S. cerevisiae Arp11 ortholog (Arp10) → defensible but coarse
GO:0005634 nucleus (is_active_in) yeast ARP9 (Q05123), Candida ARP9 (Q5A9X7, TrEMBL), mouse ACTL7A (Q9QY84), T. brucei "Actin-like protein, putative" (Q57ZL0, TrEMBL) not one is an ARP11 ortholog → wrong subfamily

The nucleus row is the clean failure. Yeast Arp9 is a subunit of the SWI/SNF
(GO:0016514, IDA) and RSC (GO:0016586, IDA) chromatin remodellers; mouse Actl7a is a
testis actin-like protein with nucleus IDA/EXP. All four sources carry their own
experimental evidence for nucleus, so the source annotations are sound and the
propagation is what fails — PROPAGATION_BAD plus WRONG_ORTHOLOG_OR_PARALOG, not
SOURCE_WEAK_OR_INFERRED. Mechanically visible in the WITH/FROM itself: dynactin
membership was propagated at node PTN000232945 and mitochondrial transport at
PTN000232947 (both ARP11-specific), whereas nucleus came from PTN008986520, a
deeper node that groups ARP11 with the nuclear ARPs.

The subfamily boundary being crossed is not my own construction. The Arps pair up by
function — Arp1 with Arp11 in dynactin, Arp7 with Arp9 in the nuclear remodellers
PMID:16291862
— and the current family review draws the same three-way split, with dynactin's Arps
cytoplasmic
PMID:42439233,
a nuclear group
PMID:42439233,
and a testis group covering ACTL7A
PMID:42439233.
Arp10p being the yeast ARP11 is likewise established, not assumed
PMID:16291862,
PMID:16291862,
which is why SGD:S000002513 on the GO:0005200 row counts as an ortholog transfer and
SGD:S000004636 on the nucleus row does not.

Independent check (RESULTS.md section D): in Human Protein Atlas
immunofluorescence, 3/4 human nuclear ARPs (ACTL6A, ACTL6B, ACTR8) show a nuclear
compartment and 0/3 dynactin ARPs (ACTR10, ACTR1A, ACTR1B) do. ACTR10's HPA locations
are vesicles, plasma membrane, focal adhesion sites and several sperm structures —
including "Perinuclear theca", which contains the string "nuclear" but is a sperm-head
cytoplasmic structure, not the nucleus. (The analysis script matches HPA's controlled
vocabulary by exact name for precisely this reason; substring matching scored ACTR10
as nuclear on the first run and inverted the result.)

6. The IPI rows

Three GO:0005515 protein binding rows, resolved from WITH/FROM:

7. The Reactome granule annotations, and the harm they have already done

Four TAS rows put ACTR10 in GO:0005576 extracellular region (×2),
GO:0035578 azurophil granule lumen and GO:1904813 ficolin-1-rich granule lumen,
via Reactome's "Exocytosis of azurophil granule lumen proteins" (R-HSA-6798751) and
"Exocytosis of ficolin-rich granule lumen proteins" (R-HSA-6800434). Reactome's
Neutrophil degranulation pathway cites bulk granule proteomics — Rørvig et al.
PMID:23650620 — as the source of these protein sets.

Reasons this is not credible for ACTR10:

  1. Topology. ACTR10 has no signal peptide and no transmembrane segment (UniProt FT
    table lists only CHAIN, sequence CONFLICTs and secondary structure), and UniProt
    gives one location: Cytoplasm, cytoskeleton. A subunit buried in a 1.2 MDa
    cytosolic complex has no route to a secretory granule lumen.

  2. Not a complex-level property — computed, not hand-checked, by
    ACTR10-bioinformatics/subunit_granule_survey.py (RESULTS.md §F), which resolves each
    reaction's pathway from Reactome rather than guessing it from the reaction id. The
    canonical roster is the 11 in UniProt's own SUBUNIT line: DCTN1/DCTN2/DCTN3 (shoulder),
    ACTR1A/ACTB (filament), CAPZA1/CAPZB (barbed end), ACTR10/DCTN4/DCTN5/DCTN6 (pointed
    end).

  3. 1/11 carries any annotation from Reactome's Neutrophil degranulation route —
    ACTR10 alone.

  4. 1/11 is annotated to azurophil granule lumen — ACTR10 alone.
  5. 3/11 carry any of the three terms by any evidence code: ACTR10, plus two that do
    not share its route — ACTB only by HDA mass spectrometry (a different
    artefact class), and CAPZA1 only from R-HSA-879377, an S100B/AGER binding
    reaction that is not under Neutrophil degranulation at all.
  6. The one protein that genuinely shares the route is ACTR1B, the β-centractin
    paralog substituting for ACTR1A in a fraction of dynactin — a twelfth protein, not
    one of the 11, and like CAPZA1 one with a large non-dynactin pool.

So the pattern is per-protein detection in granule fractions, not secretion of the
complex: if dynactin were genuinely packaged into azurophil granules, all 11 would be
there and exactly one is. (Earlier drafts of this paragraph said CAPZA1 shared the
pattern and that ACTB carried nothing. Both were wrong; §11 records how the module
found it.)
3. It has already misled a downstream paper. A 2024 HCC study describes ACTR10 as
PMID:39697424
— reading the GO annotations back out as ACTR10 biology. That is the concrete cost
of leaving them in place.

8. Actions taken

19 GOA rows, one review entry each, plus one proposed NEW row = 20 entries. (fetch-gene
seeds the two GO:0005515 IPI rows that share PMID:32296183 as a single stub; they are
split back out here so the KRT31 and DCTN4 partners are judged separately, row-for-row
with GOA.)

The five cytosol rows are all ACCEPT rather than a mix: the compartment is correct and
is where ACTR10 works, and validation requires one action per term. Each reason records
that the four later rows restate the same fact from further Reactome reactions rather
than adding evidence.

Ontology gap proposed: a new GO term for pointed-end capping of the dynactin Arp1
minifilament. The mismatch is one of scope. GO:0051694 pointed-end actin filament capping is defined as "The binding of a protein or protein complex to the pointed (or
minus) end of an actin filament, thus preventing the addition, exchange or removal of
further actin subunits", and GO:0005884 actin filament is "A filamentous structure
formed of a two-stranded helical polymer of the protein actin and associated proteins" —
whereas dynactin's backbone is, per GO:0005869's own definition, "an actin-like 40 nm
filament composed of actin-related protein". The subunits Arp11 blocks are Arp1, not
actin. GO:0005870 actin capping protein of dynactin complex is about the wrong end and the
wrong aspect regardless of status: it is a CC term for the CapZ heterodimer at the
barbed end. On status, QuickGO returns isObsolete: true, renders the label as
"obsolete actin capping protein of dynactin complex" and prefixes the definition
"OBSOLETE."; OLS adds term_replaced_by: GO:0008290 F-actin capping protein complex,
which is again barbed-end CapZ and again a CC. Separately, and recorded in the PR rather than in the GO-facing justification: the term is
still present in cache/enums/gotermenum and
cache/enums/goproteincontainingcomplexenum, so enum-membership validation can admit a
term that is obsolete upstream. Two inferences about why are withdrawn as unsupported —
see §11.

Two placements are offered in the proposal, since the choice is an editorial call about
GO:0051693's intended scope: as its child if actin filament capping is meant to cover
filaments of any actin-family protein, or under GO:0065003 protein-containing complex assembly if it is to stay strictly F-actin. (Checked all definitions against QuickGO
/complete rather than paraphrasing — an earlier draft of this justification attributed
the GO:0005884 wording to GO:0051694, which was wrong.)

Two core_functions knowledge gaps are recorded on the cargo-coupling function: the
unidentified mitochondrial outer-membrane partner, and the complete absence of any
human-system assay of the coupling.

For core_functions the schema's subunit pattern applies literally:
molecular_function: GO:0005200 structural constituent of cytoskeleton (the
subunit-specific activity) and
contributes_to_molecular_function: GO:0140660 cytoskeletal motor activator activity
(dynactin's complex-level activity, which ACTR10 does not enable on its own), with
in_complex: GO:0005869 — never locations — for complex membership.

9. Provenance / process

10. Round-2 review response (PR #2274)

Six items from ai4c-agent; none changed a GO term, evidence code or action.

Dated log — kept as written. Item 1 below records the round-2 outcome, which round 3
then falsified: subunit_granule_survey.py showed CAPZA1 does not share ACTR10's
Reactome route, so the "two of the 11" figure stated there is superseded by the 1/11 in
§7 item 2 and §11. Left unedited because it is the history of how the claim moved.

  1. Cross-subunit arithmetic, second pass. The reviewer was right that ACTR1B is a
    twelfth protein, not one of the 11 canonical subunits, so "only three of the 11" was
    3-of-12. Reframed in four YAML sites and here: two of the 11 (ACTR10, CAPZA1)
    carry a Reactome TAS granule/extracellular term, ACTR10 alone carries azurophil
    granule lumen, and ACTR1B was checked alongside as the β-centractin paralog. The
    roster is now stated explicitly from UniProt's own SUBUNIT line rather than left
    implicit, which is what allowed the count to drift twice.
  2. GO:0005870 status re-verified. It is obsolete — see §"Ontology gap" above for
    the QuickGO/OLS evidence and the cache/enums staleness that prompted the query. The
    justification no longer rests on obsoletion: the term is barbed-end CapZ and a
    cellular component, so it is the wrong end and the wrong aspect either way.
  3. BP-vs-MF framing. Correct catch: the proposed term is a BP, so it could never
    replace the GO:0005200 MF row. The reason now says so and explains why no
    proposed_replacement_terms is given.
  4. ATP vs an ADP-majority observation set. Answered with data rather than prose — new
    RESULTS.md section E surveys nucleotide-binding annotation across a 12-member
    actin/Arp panel, and section A now tallies the ligand split. Results: the Arp11 chain
    carries ADP in 14 entries and ATP in 5; only 2/12 panel members carry any term
    under GO:0000166, and both use GO:0005524 — human ACTA1 (TAS, together with
    GO:0043531 ADP binding) and yeast ACT1 (IDA). So the gap is family-wide (ACTB,
    ACTG1 and ACTR1A all carry nothing), GO:0005524 is the term GO actually uses for
    this site, and GO:0043531 or the parent GO:0032559 are named as defensible
    alternatives. Note the reviewer's premise was wrong in one detail: ACTB does not
    carry GO:0005524 (QuickGO returns zero annotations under GO:0000166 for P60709).
    The precedents are ACTA1 and yeast ACT1.
  5. The GO:0005856 MODIFY collapses into an existing row. Stated in the reason: the
    gene already has GO:0005869 as part_of by IBA, and the SubCell pipeline behind
    GO_REF:0000044 cannot emit a complex term from SL-0090, so this is a
    drop-the-redundant-ancestor recommendation, not a new located_in annotation.
  6. HPA denominators no longer shrink silently. nuclear_arp_control.py now prints
    which genes were excluded and why (ACTR5 and ACTB, both "HPA record present but no
    location reported") plus the clade sizes before exclusion, so "3/4" and "0/3" cannot
    be read as a census of the panel.

Reproducibility re-checked after all three code changes: a fresh uv run python analyze.py is diff-identical to the committed RESULTS.md.

11. Round-3 review response (PR #2274) — and two withdrawn inferences

ai4c-reviewer APPROVED, closing both round-2 🟡 items, and raised two optional 🔵
items. Both are done, because both turned out to matter more than "optional" suggested.

🔵 7 — the cross-subunit number now has a module, and it found two errors in my prose

The reviewer's point was procedural: the per-subunit Reactome check was the only
load-bearing number in the review with no module behind it, and it was the one number that
had drifted twice. subunit_granule_survey.py (RESULTS.md §F) computes it. Building it
immediately exposed two mistakes that hand-checking had missed:

  1. O15507 is an inactive (deleted) UniProt entry. I had used it as DCTN3. It returns
    no gene name and no annotations, so querying it is indistinguishable from a subunit that
    genuinely carries nothing — my "DCTN3 carries none" was vacuous. The real accession is
    O75935 (DCTN3_HUMAN). ~~The module now prints the entry name for every accession
    so this cannot recur silently.~~ Superseded — that fix does not work, see §12 🔵 3:
    UniProt follows the merge, so printing the entry name displays GFRA1_HUMAN, a
    different protein's identity, and reads as a healthy answer. The working guard asserts
    primaryAccession. Do not copy the entry-name pattern into another gene's module.
    A dead accession is the quietest possible false negative.
  2. CAPZA1's extracellular TAS is not the granule route. It comes from R-HSA-879377,
    "The TRTK-12 fragment of F-actin capping protein alpha binds the AGER ligand S100B",
    which is not under Neutrophil degranulation (R-HSA-6798695). Counting it beside
    ACTR10's, as my prose did, overstated how shared the pattern is.

With the route resolved from Reactome rather than guessed from the reaction id, the number
the four REMOVEs rest on is sharper than what I had claimed:

So the argument improved by being automated. Lesson for the campaign: the claim worth
building a module for is the one that has already drifted.

🔵 8 — the cache parenthetical is out of the GO-facing justification, and two inferences are withdrawn

Removed from proposed_new_terms[].justification, which is written for GO editors and where
repo housekeeping is noise. It lives in the PR body and here instead.

More importantly, the reviewer showed my reasoning was wrong, and checking it myself showed
a second inference was wrong too. Both withdrawn:

  1. ~~The enums are stale because GO:0005870 is absent from cache/ontologies/go.tsv.~~
    Does not follow. go.tsv is a lazily-populated label-fetch cache — fetched_date column,
    7,717 rows against GO's ~40k — so absence means "never fetched", not "removed". Its
    is_obsolete column is unreliable in the same direction: 261 rows carry an
    obsolete-prefixed label but only 26 are flagged True.
  2. ~~The enum generator's obsolete filter depends on go.tsv coverage, so a never-fetched
    term escapes it.~~ Also unsupported. The two caches have independent producers:
    cache/enums/*.csv are branch-reachability sets from the external
    linkml-term-validator/oaklib expansion (gotermenum is a flat 38,751-CURIE list),
    while go.tsv comes from this repo's own label fetches. There is no shared code path.

What is verified: GO:0005870 is obsolete upstream (QuickGO isObsolete: true; OLS
is_obsolete: true with term_replaced_by: GO:0008290) yet present in two enum caches, so
enum-membership validation can admit an upstream-obsolete term. As a corroborating pattern
only — not a mechanism — all 261 obsolete-labelled go.tsv terms appear in zero
of the 7 enum files, which makes GO:0005870 an anomaly rather than the rule. Actionable
item is a scheduled enum regeneration, not a per-PR fix.

Also noted from the reviewer, not acted on: ACTB appears under the HPA "excluded from those
denominators" line although as conventional actin it was never in either clade denominator
— the parenthetical clade label makes it readable, and the reviewer agreed it was not worth
a commit on its own.

12. Round-4 review response (PR #2274)

One 🟡 and four 🔵 from ai4c-reviewer. No GO term, evidence code, qualifier or action
changed.

🟡 The notes' standing argument still carried the claim round 3 falsified

Correct and the most important item of the round. Round 3 fixed the two YAML sites and added
§11, but left §7 item 2 — the undated, standing argument, which is where the next reader
looks — still saying CAPZA1 and ACTR1B share ACTR10's pattern, and still listing ACTB as
carrying nothing while the same paragraph's last sentence said it carries GO:0005576 by
HDA. So one file contradicted itself in three places. Rewritten from the module's numbers
(1/11 on the route, 1/11 azurophil, 3/11 by any code) with the two non-sharers named and why.
The dated round-2 log in §10 is kept as written per the reviewer, but now carries a
superseded marker pointing forward, since a log stating a falsified outcome without a
pointer misleads exactly as effectively as a stale argument.

This is the failure the brief warns about in its own words — a changed line is not a changed
claim. Appending a correction note is not the same as propagating the correction. The
generalisable fix is what was done here: after any correction, grep the whole gene folder
for the falsified claim, not just the sites the reviewer cited.

🔵 3 — the dead-accession guard, and a worse trap than I described

Round 3's claim that "the module prints the entry name so a dead accession cannot recur
silently" was wrong, and testing the guard proved it. What O15507 actually does:

This is the size=1 lesson in a new guise — an identifier lookup converting a bad input into
a confident wrong answer — and worth a campaign line of its own: a merged UniProt accession
silently substitutes another protein; check primaryAccession, never the name or the
entryType. The same guard is now in nucleotide_gap_survey.py and
nuclear_arp_control.py, which also resolve hand-written panels. paint_sources.py gets a
reported flag instead of an abort, because its accessions come from the GOA WITH/FROM field
— there a redirect is informative data about a stale GOA reference, not a code defect.

🔵 4 — no silent truncation

limit=100 with no check is the same class of defect as a denominator that silently shrinks.
All three QuickGO callers now compare numberOfHits against the returned page and abort
naming the URL. (Largest real count in this panel is 9, so nothing was truncated.)

🔵 2 — scope of the three-term screen, stated only as far as it is supported

The reviewer suggested documenting that the filter is complete because every degranulation
exocytosis reaction also emits GO:0005576. That is plausible but I could not verify it —
Reactome's containedEvents endpoint was returning HTTP 521 — so it is not asserted. The
docstring states what is supported (within this panel every route-derived set does include
GO:0005576) and names the limitation (a route reaction emitting a granule term without
GO:0005576 would be missed).

The outage also exposed a design point worth recording: route classification is a hard
dependency on Reactome, and the module aborts rather than degrading to "route unknown".
Degrading would emit a different, weaker report that still looked complete — the ABRA failure
mode. So the committed RESULTS.md cannot be regenerated while Reactome is down, and that is
the intended trade. RESULTS.md is byte-identical to the round-3 commit, and the three
modules that do not depend on Reactome — paint_sources.py, nuclear_arp_control.py and
nucleotide_gap_survey.py; all five modules require network access, so "network-independent"
would be wrong — were each re-run and confirmed to emit only lines already present in it, so
this round's code changes are output-neutral.

🔵 5 — the ficolin row's "like CAPZA1" comparison

Fixed: it now says ACTR1B and CAPZA1 are alike in having large non-dynactin pools but only
ACTR1B is on this Reactome route.

13. Round-5 review response (PR #2274) — APPROVED

ai4c-reviewer approved on f8fdcf970 and left three cosmetic suggestions, none
blocking. All three taken, because the first was the round-4 mistake about to repeat itself.

  1. §11 still carried the sentence §12 falsifies. §11 said "The module now prints the
    entry name for every accession so this cannot recur silently"     — the exact claim §12 opens
    by disproving. §10 got a superseded marker in round 4; §11 did not, and as the reviewer
    noted it is the more consequential of the two, because a reader could copy that guard
    pattern into another gene's module. Now struck through with the working alternative
    (primaryAccession) and an explicit "do not copy the entry-name pattern". Correcting a
    claim in one section while leaving it standing in another is the same failure twice; the
    rule is to grep for the falsified sentence, not the falsified section.
  2. assert_not_truncated's limit parameter was dead in both copies — the check reads
    numberOfHits against len(results) and never consulted it. Removed rather than wired
    up, since the caller's limit is already implicit in the comparison. Both copies re-tested:
    they pass a complete page and abort on a truncated one.
  3. "the three network-independent modules" was imprecise — all five modules require
    network access; the three meant are those that do not depend on Reactome
    (paint_sources.py, nuclear_arp_control.py, nucleotide_gap_survey.py). Reworded,
    since that sentence is doing the provenance work for the byte-identical RESULTS.md
    claim.

Reproducibility, resolved. Reactome's ContentService was returning HTTP 521 throughout
rounds 4 and 5, so at the close of round 5 the byte-identical RESULTS.md claim rested only on
a partial re-run: the three non-Reactome modules emitted lines already present in the file, and
section F could not be regenerated at all. A retry loop was left cycling, and Reactome came back
(HTTP 200). On attempt 9 a full end-to-end uv run python analyze.py produced a file
identical to the committed RESULTS.md — so all five modules, section F included, are now
verified reproducible, and the partial-re-run caveat above is superseded. This is the
diff-reproducibility gate met properly rather than argued around.

14. I quoted one sentence and ignored the one before it

The refuting clause was in the cached abstract the whole time. My pre-merge,
abstract-only copy of publications/PMID_12857853.md contained "found it could
coassemble" on line 49 — two lines above the sentence I quoted from line 50. A merge
conflict later replaced that file with main's full-text copy and I re-read it, which is
when I noticed; but nothing was unavailable to me when I made the error. Attributing the
catch to the merge, as an earlier draft of this section did, gets the lesson backwards and
implies the mitigation is "fetch full text" — which CLAUDE.md warns is frequently
impossible. The actual mitigation is free and always available: read the whole paragraph
around the sentence you are about to quote.

Two claims were falsified, both in the same way.

What went wrong. The abstract reads: "We tested the ability of Arp11 to interact with
conventional actin and found it could coassemble. Like Arp1, cytosolic Arp11 is found only
in dynactin, suggesting that Arp11 and free cytosolic actin do not interact
significantly." I quoted the second sentence and built on it, and the first —
sitting immediately before it and reporting the opposite result — went unused. This is
the ACBD3 lesson verbatim: quote to the end of the interpreting clause, and note that
here the disconfirming clause came before the quoted one, which is the harder direction
to catch.

The two corrections:

  1. proposed_new_terms claimed that annotating ACTR10 to GO:0051694 "would assert the
    one thing the biochemistry excludes". Wrong — Arp11 can coassemble with
    conventional actin in vitro, and its conserved barbed-end face was predicted to let it
    interact with filament pointed ends generally. The surviving argument is narrower and
    cleaner: in dynactin the subunits Arp11 blocks are Arp1, not actin, so what GO lacks is
    a term for capping the Arp1 minifilament specifically — a question of grain, not a
    prohibition on relating Arp11 to actin.
  2. The GO:0005200 reason said Arp11 "binds Arp1 but not free actin, so no
    actin-polymerisation or actin-nucleation term should follow". Overstated. The same
    paper explicitly floats an Arp2/3-like nucleation role for Arp11. It was never
    demonstrated, and the authors argue Arp1 needs no nucleator (critical concentration
    <1 nM, no lag phase) — so the right ground for withholding those terms is not
    demonstrated    , not excluded.

Both rows now cite both halves of the actin result rather than the half that suited
the argument, which is the structural fix: a one-sided quote set is the symptom.

One more datum from the full text (🔵 4). The paper maps a minimal actin-binding
fragment
PMID:12857853,
i.e. subdomains 1–2 carry Arp11's actin-binding surface. This does not conflict with the
subdomain-2 loop being deleted: the missing element is actin's 38–57 surface loop, a
sub-feature within subdomain 2, not the subdomain itself, and its absence is what stops a
further subunit adding beyond Arp11 rather than what stops Arp11 binding. Recording it because
this round's lesson is precisely about noticing the adjacent fact — and because a reader who
found this line independently might otherwise read it as contradicting §2.

Process note. I noticed the error while re-reading the file after a merge conflict
replaced it. but the file's abstract had always
contained the refuting clause, so the merge changed only when I looked, not what was there.
Nothing in the review had flagged it — the quote was verbatim, the reference title matched, and
the quote checker passed, because none of those checks can see a quotation that is true but
selectively bounded. That is the gap worth remembering: every mechanical check in this repo
validates a quote against its source, and none validates it against its own neighbouring
sentence. The reference_review for PMID:12857853 now
records the full-text reading, and the stale full_text_unavailable: true flag is
removed. (It is retained on PMID:42439233 and PMID:23650620, which really are
abstract-only.)