ACTA1 bioinformatics: WITH/FROM provenance and actin-isoform peptide specificity

Two analyses plus two consistency checks. Run each with uv run python <script> from the
repository root. The two analyses are stdlib-only; the two checks additionally use yaml
and ruamel.yaml, both already in the project environment.

Re-run order after any change: resolve_withfrom.py and actin_peptide_specificity.py
(recompute), then build_source_entities.py (rewrite), then audit_claims.py (verify).

resolve_withfrom.py --self-test exercises five guards, each stating the failure it
prevents. Two of them were written because they caught a defect in an earlier
version of this script (see "Guards" below).

1. WITH/FROM resolution: every donor resolves, and every donor is real

Counts are derived from the GOA TSV, never by hand: the script asserts that the
number of resolved sources equals the number of |-separated tokens in column 11
for every row.

26 GOA rows carry a WITH/FROM field. Zero unresolved tokens, zero unverified
mappings. Each model-organism resolution is confirmed against the resolved
UniProt entry's own cross-references, so a free-text match cannot pass as a
mapping.

The two phylogenetic rows are well founded

These are two separate numbers on purpose. The two unreviewed donors on row 1
(CGD:CAL0000191211 → A0A1D8PFR4, C. albicans ACT1; WB:WBGene00000067 →
O45815, C. elegans act-5) both carry their own IDA, but neither entry's name is
citable as evidence of what the family does, because the names are automatic.

GO:0005200 is asserted at exactly one node in PTHR11937 — PTN000940351, the
conventional-actin node — and negated by IRD at eight descendant nodes, in two
dated batches: five on 2025-08-05 (PTN000233752, PTN000233887, PTN000234048,
PTN001732543, PTN008986528) and three on 2026-04-16 (PTN000233596,
PTN000233796, PTN007551901). Its ten donors include human ACTB (EXP, IDA,
IMP, TAS), S. cerevisiae ACT1, ARP1 and ARP10, and human ARP2 and ARP3.

ACTA1 is in the retained set, checked positively rather than inferred from the
absence of a negation. A QuickGO query for GO:0005200 restricted to human and to
evidence code ECO:0000318 returns 55 annotations over 55 distinct entities (equal
here by coincidence — an annotation count is not an entity count, so the entities were
enumerated from the rows, and 55 hits fit in one page so no pagination caveat applies).
Within PTHR11937 the recipients are ACTA1, ACTA2, ACTC1, ACTG2, ACTB, ACTG1, ACTBL2,
the five POTE paralogues, ACTR10, and the divergent genes the sweep has not yet
adjudicated (ACTL8, ACTL9, ACTL10, ACTRT1-3). So this is one of the places the negation
sweep deliberately left the term standing, and the reasoning developed for the divergent
genes must not be imported here.

Four rows are self-referential, and that is valid

Rows 3-6 (GO:0001725, GO:0005865, GO:0005884, GO:0030240) have WITH/FROM
PANTHER:PTN000233075 | UniProtKB:P68133 — ACTA1 itself. The PAINT file confirms
all four as IBD at PTN000233075 seeded from UniProtKB:P68133, taxon 32523
(Tetrapoda), dated 2019-03-01. This records a PAN-GO curator judging those four
terms core for ACTA1; it is not circularity. It does mean, however, that the
quality of each row is exactly the quality of the ACTA1 experimental annotation
underneath it — which matters for GO:0001725 (see §3).

The single defective block: five ISS rows from a chicken smooth-muscle actin

Rows 19-23 (GO:0010628 positive regulation of gene expression, GO:0030027
lamellipodium, GO:0030175 filopodium, GO:0044297 cell body, GO:0090131
mesenchyme migration) all have WITH/FROM UniProtKB:P08023, which resolves to
ACTA2 — actin, aortic smooth muscle — of Gallus gallus.

The donor's evidence is sound, not weak: P08023 carries all five terms itself by
IDA/IMP, and all five trace to one paper, PMID:10633868, an antisense-knockdown
study of endothelial-mesenchymal transformation in chick cardiogenesis.

What is wrong is the propagation, on three counts:

1. Wrong paralog. Chicken ACTA2 is the ortholog of human ACTA2, not of human
   ACTA1. Chicken has its own ACTA1 (P68139, ACTS_CHICK) — the correct ortholog
   exists in the same species and was not used.
2. Wrong tissue and process. Lamellipodium, filopodium, cell body and
   mesenchyme migration are properties of a migratory mesenchymal cell.
   GO:0044297 is defined as "the portion of a cell bearing surface projections
   such as axons, dendrites, cilia, or flagella"; ACTA1 is a sarcomeric
   thin-filament protein.
3. ISS cannot discriminate here. See §2: ACTA1 is 97.9% identical to ACTA2 and
   98.9% to ACTC1, so a sequence-similarity transfer between α-actins is satisfied
   trivially and carries no isoform information.

Scope, from QuickGO. The block did not land on ACTA1 alone. It is present, by
ISS/AgBase, on all four human α-actins and on mouse Acta1:

The cytoplasmic actins ACTB and ACTG1 did not receive it. So the block was
propagated by α-actin-ness, spraying a smooth-muscle-specific developmental
finding onto the striated-muscle actins. Retracting it from ACTA1, ACTC1 and mouse
Acta1 fixes three genes in one edit.

2. Actin-isoform peptide specificity

actin_peptide_specificity.py fetches all six conventional human actins, does an
in-silico trypsin digest (cleave C-terminal to K/R, not before P; up to 2 missed
cleavages), and asks how much of ACTA1's MS-detectable peptide space is
distinguishable from the other five.

Missed cleavages are included deliberately: a peptide spanning a missed site is
more likely to carry an isoform-specific residue, so excluding them would bias the
answer towards "distinguishable".

Length guard on the comparator panel, before any scoring

A Swiss-Prot entry truncated relative to its orthologues manufactures apparent
divergence out of residues the sequence never reaches — the ACTL10 case, where a 245 aa
entry against 346-368 aa orthologues turned 20 absent positions into 20
"non-conservative substitutions". Here it would inflate the distinguishing-peptide
count, because a peptide missing from a short comparator looks unique to ACTA1. The
script therefore requires every sequence to be within 5 aa of the panel median before it
scores anything, and fails naming the offender otherwise. All six pass: 377, 377, 377,
376, 375, 375 aa against a median of 377. So the numbers below are not a truncation
artefact.

Identity

ACTA1|ACTA2 97.9%, ACTA1|ACTC1 98.9%, ACTA2|ACTC1 98.4%, ACTB|ACTG1 98.9%.
The isoforms are 375-377 aa (they differ in N-terminal processing), so 11 of the 15
pairs are unequal-length and are reported as skipped rather than compared off-frame;
the script names them.

Peptides

The nine peptides collapse to three regions, because most are nested
missed-cleavage variants of the same span. Reporting nine would overstate the
evidence:

• 3-30 DEDETTALVCDNGSGLVKAGFAGDDAPR (the N-terminus, where actins diverge)
• 287-317 CDIDIRKDLYANNVMSGGTTMYPGIADRMQK
• 338-361 KYSVWIGGSILASLSTFQQMWITK

The analysis runs on the mature chain, not the ORF — a correction from PR review.
UniProt has INIT_MET 1 "Removed", CHAIN 2..377 for the intermediate form with
N-acetylcysteine at residue 2, and CHAIN 3..377 for the mature protein after ACTMAP
cleaves that acetylated cysteine. So the ORF's N-terminal tryptic peptide — the one
beginning at Met-1, which an earlier version of this file printed — does not exist in
vivo, and a targeted-proteomics experiment built on it would order a synthetic standard
for a species that cannot be detected. (audit_claims.py treats that peptide string as
retracted and fails if it reappears on any prose surface, which is why it is described
here rather than quoted.)
Comparators contribute both their ORF and their own mature-chain digests, so a peptide is
called distinguishing only if no other actin can produce it in either form.

The counts are unchanged — 63 peptides, 9 distinguishing, 3 regions. Only the peptide
identity moves, and that is the part the experiment depends on.

What the ORF-versus-mature check asserts, and what it cannot. The two forms give the same
counts but their distinguishing sets differ in 4 members — the N-terminal peptides, which is
exactly what modelling the processing is for. So the assertion is not "the sets agree"; it
is that the counts agree and every peptide the two forms disagree about lies at the
N-terminus, a divergence anywhere else meaning the offset corrupted the digest. The
tolerance for "N-terminal" is a fixed constant grounded in the biology and deliberately not
derived from the offset under test: an earlier version used MATURE_START + 1, which let a
bogus offset of 50 buy itself a 51-residue window and pass, emitting a confident wrong region.

Break-testing that fix showed its honest limit — an offset of 4 is plausible processing and
passes — so the value is no longer hardcoded. mature_chain_start() parses it from the
CHAIN feature in ACTA1-uniprot.txt whose note matches the entry's RecName, and fails
unless that is exactly one feature. It returns 3, agreeing with the previously hand-typed
value; the point is that it is now derived from the record rather than transcribed from it.

Comparator offsets. ACTA1's mature start is 3, but that is specific to ACTA1 and is not
reused: ACTB is annotated CHAIN 1..375 and CHAIN 2..375 "N-terminally processed", so its
observable forms begin at residue 1 or 2. Every comparator therefore contributes its digest at
offsets 0, 1 and 2. This can only enlarge the comparator pool and so only shrink the
distinguishing set — conservative by construction, and it does not depend on six CHAIN
annotations being complete.

What this does and does not establish

It does not refute the five HDA rows (GO:0070062 ×3, GO:0072562,
GO:0005576), placing ACTA1 in prostatic-secretion exosomes, parotid exosomes,
trabecular-meshwork exosomes, plasma microparticles and tears. Three distinguishing
regions exist, so ACTA1-specific attribution is achievable, and a study reporting a
peptide from one of them would settle the matter.

What it establishes is that attribution is not automatic. 85.7% of ACTA1's
detectable peptide space is shared, and 29 of those peptides are shared with ACTB
and/or ACTG1 — which are expressed in every tissue those five studies sampled,
whereas none of the five sampled skeletal muscle (HPA: ACTA1 is "Tissue enriched
(skeletal)"). So these rows are untested rather than confirmed: they rest on an
isoform assignment that the analysis shows cannot be assumed.

This is the same distinction the ACTL8 review drew between its filament-interface
result (refuted) and its ATP-site result (untested) — reported separately because
the evidence differs in strength.

3. Checks run that came back negative

Recorded so the next reviewer knows they were run, not skipped.

• Reference-projection check (the ACTR8 defect). Every one of ACTA1's functional and
  localisation references was queried by PMID in QuickGO. An annotation count is not an
  entity count, so entities were enumerated from the returned rows rather than inferred
  from the total, and that is only legitimate here because each of these totals is small
  enough to fit in one page: PMID:1423520 → 1 annotation, on ACTA1;
  PMID:11333380 → 1, on ACTA1; PMID:15198992 → 3, all on ACTA1;
  PMID:12849983 → 3 — 2 on ACTA1 plus the reciprocal on DNASE1, i.e. 2 entities;
  PMID:10508519 → 5, all on ACTA1. No projection: no reference annotates a complex
  plus its subunits, and no phenotype term spreads beyond the gene assayed.

The high-count references are declared screens whose evidence codes already say so, and for
them only the annotation total was read — PMID:32814053 20,010, PMID:33961781 9,514,
PMID:28514442 3,731. Entity counts for those three are unavailable and are not
estimated from a sample: the results are paginated, so a page total is not the whole, and
substituting one would be exactly the annotation-for-entity confusion this check exists to
avoid. Nothing in the review rests on an entity count for them.
- Retraction / erratum / expression-of-concern check. 28 cited PMIDs were read
from CommentsCorrections/RefType on each cited article's own PubMed record —
not by a publication-type search, which cannot see a Publisher Correction. Zero
flagged.
- IBA-precision check (the ACRV1 defect: a propagation landing above its
donor). Rows 3-6 land at the same terms their donor holds, because the donor is
ACTA1 itself. No downward MODIFY warranted.
- Heterogeneous-donor / LCA check (the AADACL4 caveat) on GO:0005200. It does
not apply: the donor set agrees — all ten are conventional actins or Arps that
hold GO:0005200 themselves — so the term is not a broad LCA papering over a
mixed clade, and no specificity upgrade is available or needed.
- Partner-accession check (the ACRV1 TrEMBL/ORFeome substitution). All ten
annotated interaction partners resolve to reviewed entries. The one non-canonical
identifier is Q6ZQX7-4, which is a declared isoform of LIAT1 rather than a
partial clone.

4. A relayed claim, measured and refuted

An earlier draft of the GO:0043531 row placed ACTA1 among a supposed pair of family
members holding both nucleotide terms. That was a list from the merged ACTR10 review,
restated as a count of my own. nucleotide_terms_in_family.py measured it across all 533
reviewed (Swiss-Prot) PTHR11937 members:

So the claim was wrong in both directions at once: 31 reviewed members carry ATP binding
rather than two, and ACTA1 is the sole holder of ADP binding among reviewed members
rather than one of a pair. The sibling's list had been about ATP binding alone. Relay a
sibling review's claim as a claim, not as a fact — and note the corrected fact is the
stronger one.

Read the scope, because it is narrower than "the family". The member list is built from
InterPro's reviewed-only protein endpoint (fetch_interpro_family_simple.py:
/protein/reviewed/entry/...), so 533 is about 0.6% of the 88,887 proteins
PTHR11937's own metadata reports. Every figure in the table is over that reviewed subset, and
the JSON keys say so (n_reviewed_with_adp_binding,
subject_is_sole_adp_holder_among_reviewed). A count over 533 entries must not be restated as
a fact about 88,887.

Whether the result extends family-wide is an argument, not this measurement, and is
offered as one: ACTA1's GO:0043531 is a manual TAS annotation, unreviewed TrEMBL entries
receive only IEA, and no IEA pipeline maps the actin fold to ADP binding — so an unreviewed
member holding the term is unlikely. Unlikely is not measured, and the prose does not pretend
otherwise.

goUsage=descendants over is_a,part_of is set, matching the sibling resolve_withfrom.py,
so a member annotated only to a child term is counted rather than missed. Setting it left both
counts unchanged at 31 and 1.

Two traps produced a wrong answer before the right one, both worth recording:

• Querying by GO term alone paginates into a false negative. GO:0043531 has ~205,000
  annotations across GO and GO:0005524 ~9.6 million. One unpaged request returns page 1 of
  200; intersecting that with the family gives an empty set, which reads as "no member
  carries this term" — the silent-zero shape again, and it is what the first attempt returned
  for both terms. The query is therefore keyed on the family's own accessions.
• Per-accession queries do not finish (533 members × 2 terms). QuickGO accepts a batched
  geneProductId list, and each batch asserts its own result was not truncated — by comparing
  the reported hit count against the number of rows actually returned, never against the
  requested page size. That distinction is the point: a service that clamps an over-large
  limit rather than rejecting it would satisfy a constant-based check while handing back fewer
  rows, which is the silent truncation the guard exists to prevent. (Measured: QuickGO does
  honour limit=200 here — but the check must not depend on that staying true.) Re-running
  after the fix left the counts identical at 31 and 1.

5. Keeping the numbers honest

Two of the four scripts exist only to stop the review's own claims drifting, because
that - not a wrong term or a fabricated quote - is what has actually gone wrong most
often in this campaign: one claim asserted at several sites and corrected at all but one.

build_source_entities.py is the only thing that writes source_entities. Every
hand-maintained source list in this campaign has drifted, and the drift was only ever
caught by scripting a diff against the GOA. It also verifies the hand-restored
supporting_entities: the fetch-gene stub collapsed the seven PMID:32814053
interaction rows into one, and since those seven share a
(term, evidence, reference) key the test is a multiset comparison against the GOA's
WITH/FROM fields under that key - 11 interaction rows match one-to-one, no duplicates,
no omissions, no invented partners. It checks the converse direction too, so a dropped
row is an error rather than an absence.

Two things it got wrong first:

• The three-part key is not unique. The first version asserted it was and aborted on
  the seven interaction rows. The assertion was right to fire; the fix is that a key maps
  to a list of token lists, and uniqueness is demanded only where one answer is needed.
• yaml.safe_load + yaml.dump is not a round trip. It reflowed every prose scalar
  and deleted all 36 section comments - a mutation far larger than the intended one, and
  invisible to a re-run of the checker. It now uses ruamel.yaml round-trip mode and
  refuses to write if the comment count drops.

audit_claims.py also detects duplicated YAML mapping keys, which nothing else in
this repo can see: PyYAML keeps the last of a duplicated key, so a second
supported_by: under one review silently deletes the first one's entries — and the quote
checker and both validators walk the parsed document, so a quote that parsing already
removed is not there to fail. Checked two ways, a strict loader that rejects duplicates
outright and a raw-versus-parsed count of provenance entries (51 = 51).

audit_claims.py pins the numbers that appear on more than one prose surface (the review
YAML, the notes, this file) and derives every expected value from the JSON outputs, not
from literals, so a recomputation that changes a number fails the lint instead of
silently disagreeing. It also carries a list of retracted phrasings - each one a claim
that was wrong at some point in this review's history - and fails if any reappears. Its
own two false positives are worth recording: claims routinely straddle a line break in
wrapped prose (fixed by matching whitespace-normalised text), and prose legitimately
spells a small count as a word in one place and a digit in another (fixed by letting a
claim carry equivalent variants). Both were reporting a regression where none existed,
which is how a lint gets switched off.

6. Guards

resolve_withfrom.py --self-test has five cases. Three are ordinary; two exist
because they caught real defects in this script:

• Case 2 pairs a real WITH/FROM token with a real but unrelated accession and
  requires verification to fail. Without it the mapping check would be decorative.
• Case 4 revealed that the accession-match liveness guard is insufficient.
  O15507 — the dead accession the ACTR10 review found — comes back with its own
  primaryAccession and an otherwise empty record, so matching the accession
  passes it. The authoritative signal is entryType: "Inactive", and UniProt
  further reports inactiveReason: {MERGED → P56159}. The guard now reads that
  field and names the replacement. Case 5 then requires a live accession to still
  pass, so the guard discriminates rather than refusing everything.
• Case 3 covers a silent zero: QuickGO's annotation endpoint rejects every
  non-UniProtKB gene-product id used in this file's WITH/FROM fields (MGI:,
  SGD:, RGD:, WB:, dictyBase:, PomBase:, FB:, CGD: all return HTTP
  400). The script treats 400 as non-retryable, re-queries via the resolved UniProt
  accession, and records the fallback route in the output. An earlier version
  returned a vacuous 0 for the four WormBase donors, which read as "these sources
  carry no evidence" when in fact they had never been asked.