ACTL7A (Q9Y615) — review notes

Reviewer working notes for the PAINT + affinage round. Provenance is inline; every quoted
string is verbatim from the cited source.

1. What the gene is

ACTL7A ("actin-like protein 7A"; also Arp7A, T-actin 2, Tact2) belongs to a small set
of largely testis-restricted actin-related proteins, though ACTL7A itself is the least restricted
of them: HPA calls it "Tissue enriched (testis)" rather than testis-specific, and the original
human survey detected its message broadly — PMID:10373328. Mouse work states the opposite for the ortholog
PMID:21278383, so the breadth of ACTL7A expression is not settled; every functional observation is in
germ cells. It is 435 aa, intronless, and sits head-to-head with its paralogue ACTL7B on
chromosome 9q31 [PMID:10373328 "Genomic analysis revealed ACTL7A and ACTL7B to be intronless genes
contained on a common 8-kb HindIII fragment in a "head-to-head" orientation."].

Its distinguishing architectural feature is a 64–65 residue N-terminal extension in front of
the actin fold, which UniProt records as disordered (REGION 1..64 /note="Disordered") and which
Pfam models separately as PF16840 ACTL7A_N. The actin fold itself is unusually intact for an
Arp: PMID:21278383.

Note the naming trap that this review had to resolve: the family is a mixture. The PANTHER
family that ACTL7A belongs to (PTHR11937 ACTIN) also contains the nuclear Arps that are
subunits of chromatin-remodelling complexes: PMID:21278383. Several of ACTL7A's
IBA WITH/FROM sources are exactly those proteins (see §4).

2. Where it acts — acroplaxome, not acrosome

The single most consequential localisation fact for curation is that ACTL7A sits in the
subacrosomal layer (acroplaxome), i.e. between the acrosome and the nuclear envelope, and
that this was an explicit correction of the earlier "nucleus"/"acrosome" reading:

PMID:21278383

GO's GO:0033011 perinuclear theca definition already covers this region — "It shows two distinct
regions, a subacrosomal layer and, continuing caudally beyond the acrosomic system, the
postacrosomal sheath" — so GO:0033011 is the correct, if coarse, term. GO has no term for
the acroplaxome (OLS/QuickGO searches for "acroplaxome" and "subacrosomal" return nothing; the
only child of GO:0033011 is GO:0033150 cytoskeletal calyx, which is the postacrosomal part).
That asymmetry is the ontology gap proposed in this review.

Other locations, all real but secondary:

• Golgi apparatus, transiently during acrosome biogenesis: PMID:21278383.
• Nucleus of germ cells. Originally reported for mouse T-ACTIN-2 PMID:12672658, and confirmed by 3D confocal imaging in
  the knock-out study PMID:36734600. Human evidence is a sperm-nucleus proteome
  PMID:21630459 — worth noting that those preparations were made to exclude the acrosome
  ("sperm nuclei were obtained through CTAB treatment and isolated to over 99.9% purity without
  any tail fragments, acrosome or mitochondria") but the perinuclear theca is nuclear-adherent by
  construction, so PT carry-over cannot be excluded from a nuclear proteome alone. The mouse IF
  data are what make the nuclear pool credible. No nuclear function has been demonstrated.

3. What it does — and the crucial thing it does not do

Mouse knock-out is the decisive experiment: PMID:36734600 and, importantly, the effect is local:
PMID:36734600

Whether ACTL7A is itself the filament is explicitly open: PMID:36734600 The 2011 paper closes with the same open question: PMID:21278383

And there is a direct, if unillustrated, cell-biological negative on self-polymerisation:
PMID:21278383

Partners (all in the perinuclear theca / acroplaxome orbit):

• TES (testin) — direct, structurally defined. UniProt: Interacts (via N-terminus) with TES (via LIM domain 2). The crystal structure 2XQN is of ACTL7A 1–65 bound to TES LIM2-3 with the
  ENAH EVH1 domain PMID:21278383,
  and single alanine-to-tyrosine substitutions in ACTL7A abolish it (MUTAGEN 31 A->Y: Abolishes interaction with TES.; same for residue 41). This is a LIM-domain-binding activity, not
  generic protein binding.
• ACTL9 PMID:33626338, CYLC1/cylicin-1 PMID:38573307, both PT proteins whose own loss
  causes PT/acrosome-attachment defects.
• ZPBP, ARP2, MYO6, PFN4, ACTN1, DCTN1 by co-IP from the KO study; ACTRT2 was tested and was
  negative PMID:36734600.
• FNDC8 and CCIN, in mouse PT PMID:41169243.

4. WITH/FROM resolution and source evidence (the high-yield check)

Every WITH/FROM token in ACTL7A-goa.tsv, resolved, with what QuickGO says the source itself
carries for the propagated term.

Two things fall out of this table.

1. The nucleus IBA is not a fold artefact, even though the family is mixed. I expected the
   ARP9 sources to be the whole story; they are not. The mouse ortholog carries its own IDA/EXP
   nucleus evidence, and even a specific term (GO:0001673 male germ cell nucleus) from the same
   1999–2003 T-actin work. So the term is right; what is not established is any nuclear
   activity, which is what is_active_in asserts.
2. The structural molecule activity IBA has no protein source — but its tree nodes turned out
   to be the most informative sources in the whole review. Both WITH/FROM tokens are internal
   PANTHER nodes, and across the ACTL7 orthologue set (human ACTL7A, mouse Actl7a/Actl7b, rat
   Actl7a) the only non-IBA support for GO:0005198 or any descendant is the human gene's own
   GO:0005200 TAS row, citing the 1999 cloning paper. It was easy to stop at "not a protein,
   therefore carries nothing" — an earlier draft of these notes recorded both tokens' evidence as
   n/a, and in doing so dropped the fact that decided the molecular function. Opening the cached
   PAINT table for the family shows those nodes encode a deliberate curatorial decision:
   GO:0005200 is IBD-supported at the ancestral node from real actins and negated by IRD on
   descent to the node carrying ACTL7A, with GO:0005198 substituted the same day. See §9.
   An unresolvable WITH/FROM token is not the same as an uninformative one.

5. The 1999 TAS row is a name-derived annotation

GO:0005200 structural constituent of cytoskeleton / TAS / PMID:10373328 (source: ProtInc).
The cited paper is a positional-cloning paper: cDNA selection, genomic sequencing, linkage
mapping, and Northern expression analysis. Its abstract contains no functional or biochemical
experiment; the only structural statement is homology-based ("Two novel human actin-like genes,
ACTL7A and ACTL7B, were identified by cDNA selection and direct genomic sequencing"). The
"structural constituent of cytoskeleton" claim therefore comes from the name "actin-like", not
from data — the exact failure mode this campaign keeps finding, where a fold or domain name is
promoted to an activity.

It then propagates: GO:0007010 cytoskeleton organization is attached by GO_REF:0000108, whose
WITH/FROM is literally GO:0005200 — an inter-ontology MF→BP link. So one 1999 name-based
inference produces both an MF and a BP row.

Notably, the conclusion is largely defensible on 2011–2025 evidence (the acroplaxome and PT are
cytoskeletal structures — GO:0033011 is itself a descendant of GO:0005856 cytoskeleton — and the
KO does lose subacrosomal F-actin). But the cited evidence does not support it, and the BP that
falls out of it is far less specific than what the mouse KO actually shows.

6. Structure-guided audit of the fold (this review's own analysis)

Full method and tables: ACTL7A-bioinformatics/RESULTS.md, regenerated by
uv run --script actin_fold_audit.py. Contact residues are computed by neighbour search from
PDB 2BTF (profilin–β-actin, ATP) and PDB 8A2S (cryo-EM F-actin, Mg-ADP-Pi, 5 protomers)
and mapped through a MAFFT alignment; nothing is hard-coded, and every position is shown in a
gap-free alignment window so the mapping can be checked.

Three results matter for curation.

• The nucleotide cleft is largely retained. ACTL7A is 63.2% identical to actin across the 19
  positions within 4.0 Å of ATP or the divalent cation (66.7% at the phosphate contacts). The
  calibration: conventional actins 96.5%, Arp2/Arp3 79.0%, and the genuinely divergent yeast
  SWI/SNF Arps (Arp7/Arp9) 36.9%. ACTL7A is nowhere near the "lost it" end of the family.
  [file:human/ACTL7A/ACTL7A-bioinformatics/RESULTS.md " Nucleotide cleft largely retained.* ACTL7A is 63.2% identical to actin across the 19 G-actin cleft positions (66.7% at the phosphate contacts), against 96.5% for conventional actins, 79.0% for Arp2/Arp3 and 36.9% for the divergent SWI/SNF Arps."]
• The ATP-hydrolysis trigger is not retained. Of the five literature-defined catalytic
  positions (Asp11, Gln137, Asp154, Val159, His161 — PMID:37009486, PMID:30622175) ACTL7A keeps
  three: DQEVY. Asp154→Glu is conservative; His161→Tyr is not, and His161 is precisely the
  residue whose side-chain flip on filament incorporation triggers hydrolysis. Every conventional
  actin and Arp1/Arp2/Arp3 in the panel keeps all five. ACTL7B keeps His161; ACTL7A and ACTL9 do
  not. The alignment here is anchored on an identical PIYEGY motif on the C-terminal side, so
  this is not an alignment artefact.
• The filament interface is not retained. Across the 79 inter-protomer contact positions of
  F-actin, ACTL7A matches actin at 42.3%, versus 93.2% for conventional actins and 55.7% for
  Arp1, which does form a filament (in dynactin). This is quantitatively consistent with the
  cell-biological negative in PMID:21278383.

And a fourth result that guards against the opposite error — writing the pocket off as vestigial:
reported patient variants cluster in the cleft. 20 of ACTL7A's 435 residues (4.6%) align to a
cleft position, yet 2 of 5 variants reported in SPGF86 patients land there (G246A ↔ actin
Gly182; G362R ↔ actin Gly301, a phosphate contact), against 0 of 4 population polymorphisms;
exact binomial p = 0.019. A third, D75A (PMID:36574082), maps to actin Asp11, part of the
divalent-cation site, just outside the 4.0 Å shell. The sample is small and biased, and UniProt
flags A245T/G246A/G362R as "uncertain significance" — but the pocket is clearly not decorative.

Synthesis. ACTL7A is not a canonical actin and not an emptied-out fold. It most likely still
binds nucleotide (pocket conserved, disease variants concentrated there) while being unable to
run actin's hydrolysis cycle (His161 lost) and unable to build a canonical filament (interface
degenerate, plus the direct negative in cells). That is exactly the profile of a structural /
scaffolding actin-fold protein, which is what the acroplaxome phenotype demands. Nothing here
justifies annotating polymerisation; nothing here justifies asserting the pocket is dead either.
No ATP-binding assay on ACTL7A exists, so no MF annotation is proposed — it goes in
suggested_experiments.

6b. ACTL7A versus ACTL7B — the paralogue pair, checked rather than assumed

ACTL7A and ACTL7B are head-to-head neighbours with >65% identity to each other
PMID:10373328, which makes them the obvious candidates for annotation cross-transfer. Mouse
Actl7b (MGI:MGI:1343053) is in fact already a WITH/FROM donor on ACTL7A's GO:0005737
cytoplasm IBA row. So the pair was checked on every axis where this review makes a claim, rather
than assumed to behave alike. They are separable on all four.

ACTL7B has since been reviewed independently and merged, so the comparison below is against that
review (genes/human/ACTL7B/ACTL7B-ai-review.yaml) as well as against the primary literature.

Two corrections this cross-check forced on an earlier draft of these notes, both from overstating the
contrast. ACTL7B is not purely cytoplasmic — the merged review gives it a low-abundance nuclear
pool in early spermatids and spermatocytes, so PMID:12672658's nucleus-versus-cytoplasm split is a
statement about the dominant pool in haploid germ cells, not an absolute presence/absence
difference. And "not in sperm" is specifically about mature sperm; ACTL7B is on the developing
acrosome during spermiogenesis.

Four consequences.

1. The GO:0005737 cytoplasm IBA rests partly on the paralogue's half of the experiment.
   PMID:12672658's localisation sentence — PMID:12672658 — is only usable with the naming key: PMID:12672658 T-ACTIN-1 is ACTL7B.
   So that paper's cytoplasm evidence belongs to the paralogue and its nucleus evidence to ACTL7A.
   The cytoplasm transfer survives only because mouse Actl7a has independent cytoplasm IDAs
   (PMID:21278383, PMID:36734600). The donor is now marked SUPPORTS_SOURCE_BUT_NOT_TARGET.
2. GO:0030274 LIM domain binding must not be assumed to transfer to ACTL7B. The TES
   interaction is carried entirely by ACTL7A's N-terminal extension, and that extension differs
   between the two: PMID:21278383
   Both proteins reach the acroplaxome; only ACTL7A is known to get there partly via TES.
3. The His161 split is the sharpest single discriminator in the family. ACTL7B keeps actin's
   hydrolysis trigger; ACTL7A and ACTL9 replace it with tyrosine. That is a within-paralogue
   functional difference, not shared drift, and it means a "cannot run actin's hydrolysis cycle"
   argument made for ACTL7A does not carry over to ACTL7B. Note that the two independent
   analyses agree on the filament conclusion by different routes: the ACTL7B review finds "the
   surface actin uses for filament formation is not conserved, with the DNase-I-binding loop
   entirely replaced", and this review's contact-based audit puts ACTL7B's protomer interface at
   36.4% — the lowest of the two paralogues.
4. The two reviews initially reached opposite verdicts on GO:0005200; they now agree, and how
   that resolved is the finding. ACTL7B's review removed the GO:0005200 TAS row; drafts 1-2
   of this review kept it, and draft 2 promoted it to the core MF, arguing the pair could diverge
   because ACTL7A's GO:0033011 is a GO:0005856 descendant while ACTL7B's GO:0001669 is
   not. That ontological asymmetry is real (verified via QuickGO ancestors on both terms) but it is
   not sufficient, because residing in a cytoskeletal structure is not being a structural
   constituent of one. What settled it is a record neither the asymmetry argument nor the sibling
   review's reasoning needed: GO's own pipeline has explicitly negated GO:0005200 for the
   PANTHER node that carries ACTL7A, dated after all the knock-out and patient literature (§9).
   Both reviews now remove the row, for the same reasons. The general lesson survives intact and is
   in fact sharpened: a name-derived annotation is applied per family, so when it is wrong it is
   wrong for the whole family — and GO's phylogenetic pipeline had already worked that out, in a
   file sitting in this repository.

One further cross-gene note. ACTL7B's review adds GO:0001669 acrosomal vesicle as a NEW
annotation on the strength of the protein being on the developing acrosome, whereas this review
argues the same term away from ACTL7A because a cytosolic protein cannot be inside the
hydrolase-containing vesicle. These are not necessarily in conflict — different evidence, different
gene — but "on the developing acrosome" is exactly the description that the missing acroplaxome
term would disambiguate. The proposed term therefore has at least two immediate users in this
paralogue pair alone, which strengthens the case for it.

7. Human disease genetics, and the GO gap it exposes

Bi-allelic ACTL7A variants cause spermatogenic failure 86 (SPGF86, MIM:620499), presenting as
normal-looking semen with total fertilization failure or early embryonic arrest after IVF/ICSI
[PMID:32923619; PMID:34727571; PMID:36593593; PMID:37004249]. The mechanistic through-line is
oocyte activation: PMID:34727571 and the defect is rescued pharmacologically: PMID:34727571 The same axis is reported in mouse PMID:32923619 and for the G402S knock-in
PMID:35863052.

GOA's most specific process term for ACTL7A is GO:0009566 fertilization. GO:0007343 egg activation exists and is a descendant of GO:0009566, and human PLCZ1 carries it as
IMP PMID:26721930 — i.e. GO already uses this term for a sperm-borne factor whose loss causes
oocyte-activation failure in patients. ACTL7A is the same case and does not have it. That is the
main annotation gap proposed here.

8. Missing process term for the F-actin phenotype

Neither human ACTL7A nor mouse Actl7a has any annotation to GO:0030036 actin cytoskeleton organization, GO:0007015 actin filament organization, or any descendant (checked by QuickGO
goUsage=descendants on Q9QY84 and Q9Y615) — even though the central result of PMID:36734600 is
total loss of subacrosomal F-actin in the KO. The only actin-process row present is the
inter-ontology GO:0007010 cytoskeleton organization, which arrived via the 1999 TAS MF. Moving
that row to GO:0030036 puts the correct, evidenced process in place of a parent that was reached
by the wrong route.

9. Decisions and their rationale

How the two structural MF rows were resolved (settled in round 5, after two wrong turns). This
was the hardest call in the review and it took three attempts, so the reasoning is recorded in full.

Draft 1 flagged GO:0005200 as over-annotated on provenance grounds while ACCEPTing its vaguer
parent GO:0005198 and promoting that to the core molecular function — incoherent, because the
draft's own reason argued GO:0005200's term was defensible and needed only re-sourcing, and
core_functions[0].description began "Structural constituent of the acroplaxome".

Draft 2 over-corrected: it ACCEPTed GO:0005200, made it the core MF, and argued that the
paralogues could legitimately diverge because ACTL7A is annotated to GO:0033011 (a GO:0005856
descendant) while ACTL7B's GO:0001669 is not. It also dismissed the possibility that PAINT had
declined the child deliberately, on the grounds that a family-level generalisation "cannot see a
gene-specific knock-out published years later".

That second argument was factually wrong, and the record is in this repo.
interpro/panther/PTHR11937/PTHR11937-paint.tsv shows GO's pipeline did not omit the child — it
negated it, for the node that carries ACTL7A, with the parent substituted on the same date:

PTHR11937  PTN000940351  GO:0005200  F  IBD  false  SGD:...|UniProtKB:P61160|...|UniProtKB:P60709|...  20250805
PTHR11937  PTN008986528  GO:0005198  F  IBA  false  PANTHER:PTN000940351   20250805
PTHR11937  PTN008986528  GO:0005200  F  IRD  true   PANTHER:PTN000940351   20250805

So GO:0005200 is IBD-supported at the ancestral actin node from conventional actins and
Arp2/Arp3, and rejected on descent into the divergent-ARP branches (8 IRD rows, PTN008986528 among
them). GO:0005198 appears nowhere else in the family table — it exists at this node only as
the deliberate generalisation that replaced the rejected child. Q9Y615 is in
PTHR11937-entries.csv, so the rejection covers ACTL7A directly. Dates vary per row in that file
(20190301 … 20260528), so 2025-08-05 is a per-annotation date, and it postdates the 2023 knock-out
and the 2020–2021 patient papers. The "cannot see" rebuttal therefore fails on dates as well as on
substance.

Draft 3, final. Defer to the IRD, which is also where this review's own bioinformatics points:
- GO:0005200 → REMOVE. Three converging arguments, each independent: the cited reference
contains no functional data of any kind (TAS from a gene name, so removal overrules no curator
who read data); GO's pipeline has explicitly and currently negated the term for this clade; and
the filament interface by which real actins earn the term is not conserved (42.3% vs 93–100%).
- GO:0005198 → ACCEPT, and it is the core MF. Not a fallback but the level GO chose on
purpose, and the level the phenotypes actually support — "structural integrity of a complex" is
what a knock-out that loses subacrosomal F-actin and sheds its acrosome demonstrates.
- core_functions[0] reworded from "Structural constituent of the acroplaxome" to "Architectural
component", and it now says why the weaker MF is the right one.

The withdrawn argument is left on the record rather than deleted: residing in a cytoskeletal
structure is not the same as being a structural constituent of it, and that distinction is the whole
of the mistake. This also brings the pair into agreement — ACTL7B's review removed the same row on
the same reasoning.

Why the three GO:0001669 rows get the same MODIFY, including the two experimental ones. The
argument is definitional rather than a challenge to anyone's data. GO:0001669 is "A structure in
the head of a spermatozoon that contains acid hydrolases ... derived from the lysosome", with
exact synonyms "acrosome" and "acrosomal granule" — a membrane-bounded vesicle. ACTL7A is a
cytosolic cytoskeletal protein sitting on the nuclear side of the inner acrosomal membrane, so
located_in that vesicle is a category error regardless of how good the immunofluorescence is;
what the IF shows is an acrosomal-region signal, which at light-microscope resolution
necessarily includes the acroplaxome underneath. The same paper that reports the human
G402S phenotype describes wild-type ACTL7A as attached to the acroplaxome. MODIFY (not
REMOVE) is the right instrument: the observation is real and the intended compartment is one
step away. Both reasons say explicitly that a curator with figure access should confirm.

Why no REMOVE anywhere. The two structural-MF rows were the other candidates, and neither turned
out to need it — see the paragraph above on how they were resolved. In both cases the term
survives scrutiny even though its evidence does not, and the instruments that fit are ACCEPT
with the provenance criticism inside reason (GO:0005200) and KEEP_AS_NON_CORE for a redundant
parent (GO:0005198). Nothing on this gene overshoots its term, so MARK_AS_OVER_ANNOTATED is
unused; the defects were the citation and the inference chain.

Why the nucleus rows are treated differently from each other. Only the ISS row is refined to
GO:0001673 male germ cell nucleus, and it is refined because its sole donor, mouse Actl7a,
already carries that specific term by IDA. A propagated annotation that is less precise than the
evidence its own donor holds throws information away for nothing, and refining downward is the fix —
the same pattern found on ACRV1, where a mouse donor's specific acrosomal IDAs propagated three
levels up. The other three rows are KEEP_AS_NON_CORE for reasons specific to each: the
phylogenetic row's donor set spans a yeast SWI/SNF subunit, a Candida homologue and a
trypanosome protein, for which a germ-cell term is meaningless, so GO:0005634 genuinely is the
correct least common ancestor there — a broad term can be right precisely because its donors are
heterogeneous; the GO_REF:0000120 row's WITH/FROM mixes the ortholog with a generic
subcellular-location keyword; and the HDA row is a human experimental observation best left as
the curator recorded it. This produces a per-term action split and a validator warning, which is
accepted rather than smoothed over, because the routes genuinely differ and the row-level reasons say
so. Note what is not a reason: the 1999 Northern survey. An mRNA tissue distribution says nothing
about where the protein sits, and citing it as a hedge was a first-draft error corrected in round 2.
The substantive point about the nucleus remains that no nuclear activity has been demonstrated,
which is why nothing nuclear appears in core_functions.

Two evidence-code observations, recorded here rather than in row summaries. GO:0001669 /
IMP / PMID:34727571 uses an IMP code for a location; the underlying data are
immunofluorescence on patient sperm, which is an IDA-shaped observation. And the mouse-derived
rows are correctly ISS/IEA rather than IDA/IMP — the mouse knock-out and knock-in phenotypes
justify ortholog transfer to the human gene, not direct-assay codes on it.

10. Affinage deep-research record

self_evaluation_pairwise: win (trust gates clear at fetch time), faith 83.3%, 14 citations. The narrative is broadly accurate and its PLCζ
framing is confirmed against the primary papers. Two cautions applied here:

• One citation, PMID:bio_10.1101_2025.03.27.645694, is a bioRxiv DOI in a PMID-shaped field,
  not a PubMed record; the ACTRT1/ACTRT2/ARPM1/ZPBP co-IP claim resting on it is excluded from
  this review.
• PMID:38464253 is a preprint (bioRxiv); its HDAC1/HDAC3 and INO80/SWI-SNF HSA-domain-docking
  claims are interesting but in-silico/preliminary and are not used to support any annotation.
  They are the reason the nuclear pool is flagged as a knowledge gap rather than dismissed.

Its own GO grounding lists GO:0008092 cytoskeletal protein binding and GO:0005198 structural molecule activity as the molecular activities and GO:0005634 nucleus, GO:0005856 cytoskeleton
as localisation — coarse, and it misses the acroplaxome/PT entirely, so it was not imported.

11. Review round log

Round 2 (PR #2271, ai4c-agent CHANGES_REQUESTED). Two IMPORTANT items, both correct, both
fixed:

1. A supporting quote that argued the opposite. The GO:0005737 cytoplasm IBA row cited
   PMID:12672658's "T-ACTIN-1 was found in the cytoplasm and T-ACTIN-2 was located in the nucleus".
   Since T-ACTIN-2 is mACTL7A, that sentence is evidence for the nucleus on this gene and for
   the cytoplasm on the paralogue. Replaced with PMID:21278383 and PMID:36734600 quotes that
   actually bear on ACTL7A, the naming key added as a reference finding, and the paralogue donor
   MGI:MGI:1343053 downgraded to SUPPORTS_SOURCE_BUT_NOT_TARGET. The general lesson: in a paper
   that studies a paralogue pair under aliases, the alias-to-gene key is part of the evidence and
   has to be quoted alongside the claim.
2. Flagging the specific MF down while promoting its vague parent. Resolved as described in §9.

Non-blocking items also taken: core_functions[2] now says why it asserts no molecular function
rather than leaving the slot silently empty; the GO:0001669 IMP row records that UNDECIDED was
considered and why MODIFY was preferred; and the GO:0001673 deferral no longer leans on an mRNA
Northern survey to hedge a protein-localisation question — the real reasons are donor heterogeneity
on the phylogenetic row and the absence of any somatic-nucleus experiment either way.

Also added this round, at the coordinator's prompt, the systematic ACTL7A/ACTL7B comparison in §6b.
That was worth doing independently of the review: it turned up the paralogue-donor problem on the
cytoplasm row from the other direction, and it establishes that GO:0030274 LIM domain binding
should not be assumed to transfer to ACTL7B, whose N-terminal extension is a different
sequence.

Round 3. One stale machine-readable field: the GO:0005198 row had become
KEEP_AS_NON_CORE while its propagation_review.root_cause still read NO_FAILURE_CORE, i.e. the
exact claim round 2 retracted, in the field an aggregator would read. Fixed, and then checked
mechanically across all 31 rows rather than by eye, since no validation rule cross-checks action
against root_cause. Two notes paragraphs in §9 that still argued the superseded decisions were
rewritten (see above).

Round 4. Approved by the reviewer at the end of round 3; this round takes the two remaining
non-blocking items and acts on a cross-gene comparison that only became possible once the paralogue
was merged.

• SOURCE_EVIDENCE_WEAK added to the GO:0005198 row's failure_modes. Its other propagation
  defect — both WITH/FROM tokens being PANTHER internal nodes with no protein-level evidence — was
  argued in prose but absent from the machine-readable field, which is the same aggregator argument
  accepted for root_cause a round earlier.
• The ISS nucleus row refined to GO:0001673 (see §9), applying the ACRV1 rule that an IBA or
  ISS can be less precise than its own donor.
• §6b rewritten against the now-merged ACTL7B review rather than against the primary literature
  alone. This corrected two overstatements in the previous draft — ACTL7B does have a nuclear pool,
  and "not in sperm" is about mature sperm only — and surfaced the fact that the two reviews reach
  opposite verdicts on GO:0005200. That divergence is now argued explicitly on both the
  GO:0005200 and GO:0005198 rows, resting on a checkable ontological asymmetry (GO:0033011 is
  a GO:0005856 descendant; GO:0001669 is not) rather than on either reviewer's preference.
• The core_functions[0] description now states that the constituency claim is about the acroplaxome
  plate and not about the actin filaments inside it, so it no longer appears to presuppose the answer
  to its own knowledge gap.

Round 5. The reviewer found the thing that made two earlier rounds wrong, and it was cached in
this repository the whole time: interpro/panther/PTHR11937/PTHR11937-paint.tsv records GO's
pipeline negating GO:0005200 for the node that carries ACTL7A, not merely omitting it, dated
after the knock-out and patient literature. §9 documents the three-draft path to the settled
position. Consequences: GO:0005200 TAS → REMOVE; GO:0005198 IBA → ACCEPT and core MF;
core_functions[0] reworded to "Architectural component"; the two PANTHER nodes moved off
NOT_RELEVANT to SUPPORTS_TRANSFER with the IBA/IRD pair recorded on each; the PAINT table added
as a cited reference with the two rows as verbatim findings; §4's table corrected, since it had
recorded both tokens' evidence as n/a; and the round-4 SOURCE_EVIDENCE_WEAK failure mode dropped,
because a deliberate curatorial generalisation seeded from an IBD over real actins is not weak
evidence. Two of my own arguments are withdrawn on the record rather than deleted: that PAINT "cannot
see" a later knock-out (false on dates), and that residence in a GO:0005856 descendant establishes
constituency of it (a category error).

The durable methodological lesson, and the reason this took five rounds: "PANTHER:PTN… is a tree
node, not a protein" is a true statement that invites you to stop looking. The campaign rule says
an unresolved WITH/FROM cannot be dismissed, only deferred — the same applies to an unresolvable
one. A PANTHER node cannot be resolved to a protein, but it can be looked up in the family's PAINT
table, where it may carry an explicit negation of exactly the term under discussion.