BAIAP2 (IRSp53) — curation notes

UniProt Q9UQB8. Human. GOA snapshot refreshed during this pass (see "Stale snapshot"
below); all claims cross-checked live against QuickGO.

What the protein does

IRSp53 is a curvature-sensitive membrane–actin adaptor. Three modules, three jobs:

In neurons the same adaptor logic operates postsynaptically, via PSD-95/PSD-93 and the
Shank family
PMID:15673667.

A point worth carrying forward, because it cuts against the obvious reading of the
bundling annotations: Mattila et al. found the filopodia-forming activity of the IMD to
depend on membrane deformation rather than on F-actin bundling or GTPase binding
PMID:17371834.
Actin binding is a real property of the domain but not the mechanism by which protrusions
are built, and core_functions orders the four entries accordingly.

Audit of the prior review (this pass)

The review was already status: COMPLETE from the PAINT no-IBA batch (PR #2956) and
passed validation. Four defect classes were found. All are fixed; the file went from 118
to 131 annotations and from 2 to 4 core functions.

1. Fabricated citations that passed validation via an escape hatch

Seven supported_by entries cited bare DOIs (DOI:10.7554/eLife.72316,
DOI:10.1242/jcs.262064) that appear nowhere in the references list. Six of the seven
quotes are verbatim text from BAIAP2-deep-research-falcon.md, attributed to primary
papers they were never taken from; one ("IRSp53 recognizes ~100 nm PM evaginations…") is
a paraphrase occurring in no source at all. One still carried its bullet label,
"Sites of action: …".

Why they passed is worth recording precisely, because the obvious explanation is wrong.
It is not that a DOI citation is unverifiable. publications/ holds 232 DOI-keyed records,
validation/supporting_text.py resolves DOI:x/y to publications/DOI_x_y.md, and both
DOIs here were already cached — DOI_10.7554_eLife.72316.md is 83 KB of full text,
committed in the same PR (#2956) that wrote these quotes. The verbatim check also does
cover annotation-level supported_by, not just references[].findings[].

The actual reason is that each of those entries carried full_text_unavailable: true, and
that flag suppresses the substring check outright — irrespective of reference type, and
irrespective of whether the publication is cached with full text. Verified by injecting
"PURPLE ELEPHANTS CATALYSE THE RIBOSOME ON TUESDAYS." as the supporting_text on
PMID:37747150 (83 KB of cached full text): caught as Text part not found as substring
without the flag, ✓ Valid with it. Five of the six original quotes are likewise genuinely
absent from the cached DOI files, so the fabrication finding stands; the flag is what let
them through.

The flag therefore does double duty — a legitimate signal on abstract-only records, and a
blanket verification opt-out — and the two uses are indistinguishable in the file. Anyone
auditing a review here should treat full_text_unavailable: true on a supported_by entry
as "this quote was never checked", not as a property of the cited paper, and should confirm
the flag against the cached record rather than trusting it.

Three of the seven sat on rows GOA has since retired and went away with them. The remaining
four were re-grounded on primary literature with quotes verified verbatim against the
cache: GO:0030838 and GO:0007009 → PMID:37747150, GO:0043197 → PMID:24639075, GO:0061003 →
PMID:15673667.

The DOI:10.1242/jcs.262064 citation was doubly wrong: it resolves to PMID:39404604,
"HIV-1 assembly — when virology meets biophysics", cited in support of adaptor activity.

Where this pass sets full_text_unavailable: true itself (on the PMID:19366662 reference,
whose cached record really is abstract-only and about the paralog), it is set on the
reference entry and the unverifiable quotes were deleted rather than left flagged — so no
supporting_text in this file is exempted from checking.

2. Six IEA calls dismissed on rationales that misstate their provenance

Every MARK_AS_OVER_ANNOTATED row was justified with some version of "not supported" or
"may be transferred from ortholog data". Tracing each GO_REF:0000107 Ensembl transfer
back to its source shows all six rest on a rodent experimental annotation by RGD or
SynGO — curators who read the full text. Per the CLAUDE.md rule against overruling
curators from incomplete evidence, all six were re-decided, and the review now contains
no MARK_AS_OVER_ANNOTATED calls at all.

Term Real source Was Now
GO:0001221 transcription coregulator binding rat Baiap2 IPI, PMID:10332026 (RGD) OVER_ANNOTATED KEEP_AS_NON_CORE
GO:0005874 microtubule rat Baiap2 IDA, PMID:24639075 (RGD) OVER_ANNOTATED KEEP_AS_NON_CORE
GO:0030141 secretory granule rat Baiap2 IDA, PMID:17120053 (RGD) OVER_ANNOTATED KEEP_AS_NON_CORE
GO:0098685 Schaffer collateral - CA1 synapse mouse Baiap2 IDA+IMP, PMID:19193906 (SynGO) OVER_ANNOTATED ACCEPT, then dropped — GOA retired the row
GO:0099523 presynaptic cytosol rat Baiap2 IDA, PMID:15673667 (SynGO) OVER_ANNOTATED KEEP_AS_NON_CORE
GO:0099524 postsynaptic cytosol rat Baiap2 IDA, PMID:15673667 (SynGO) OVER_ANNOTATED ACCEPT

Three deserve comment.

GO:0001221 was called unsupported, but its source is PMID:10332026, the DRPLA paper
showing atrophin-1 binding the IRSp53 SH3 domain. Atrophin-1 is a transcriptional
corepressor, so GO:0001221 is the correctly-typed molecular function — and since human
GOA has now retired its own generic GO:0005515 row citing that paper, this transferred
term is the only remaining record of the interaction.

GO:0099523/GO:0099524 were dismissed as "overly specific; cytosol annotation is
sufficient". Specificity at that granularity is the purpose of the SynGO term set, and
both are SynGO IDAs.

GO:0005874 microtubule is the one where the verdict survives in spirit but the reasoning
did not: it is a genuine rat immunogold EM observation
PMID:24639075,
not an ortholog artefact — but proximity is not function, so it is non-core rather than
dismissed.

3. Two NEW annotations cited a generated summary as IDA evidence

GO:0140090 and GO:0005546 were proposed with
original_reference_id: file:human/BAIAP2/BAIAP2-deep-research-falcon.md, and the first
rationale referred only to "eLife 2023" with no identifier. A generated summary is not
admissible evidence for an IDA. Re-grounded:

Comparator check on GO:0140090: the term is already carried in human by the
BAR-superfamily proteins ARFIP2, ICA1 and PICK1, each IDA from PMID:29768204, so
annotating curvature sensing to a BAR-domain protein is established practice and IRSp53's
absence is a gap. The activity is executed by IRSp53's own I-BAR domain, so it clears the
"which entity performs the step" test.

4. proposed_replacement_terms invented from bulk interaction data

Thirteen REMOVE rows sourced from high-throughput interactome screens proposed a
specific molecular function (GO:0030674 or GO:0097110) inferred from the fact of a
binding alone. Naming a function from an interaction list is exactly what the skill
forbids; the replacements were withdrawn and each row now says why. Nothing is lost —
both terms are independently annotated and accepted elsewhere in the review.

The mirror error was also present: GO:0003779 actin binding was proposed as a
replacement on a row whose evidence is an ACTB co-immunoprecipitation (PMID:19171758).
An adaptor pulls down whatever its partners are attached to; that is not actin-binding
evidence. Withdrawn there and asserted once on its real in vitro source (below).

5. A missing molecular function

GO:0003779 actin binding was added as NEW and as a fourth core function. BAIAP2
carries GO:0051017 and GO:0051764 as processes but had no actin-binding molecular
function anywhere in GOA, so the activity producing those processes was unrepresented.
The evidence is in vitro on purified IMD and purified actin (PMID:14752106), with the
residues mapped
PMID:15635447.
Comparator: MTSS1 carries GO:0003779 by IDA (PMID:12570871) and IBA, MTSS2 by IEA and
IBA, so there is no convention against the family. Their IBA sits on node PTN004481656,
not BAIAP2's PTN001022094, which is why it has not propagated — the argument here is
target-specific direct evidence, not a propagation claim.

GO:0180020 membrane bending activity was considered for the deformation half of the
I-BAR activity and rejected: in human it is carried only by CHMP2A, CHMP3 and OPA1,
by no BAR-superfamily protein at all. GO:0097753 membrane bending (a process, not a
function) is similarly sparse. That systematic absence reads as a convention not yet
identified rather than a gap to fill, so the point is raised in suggested_questions
rather than asserted.

6. Stale GOA snapshot

The cached tsv predated a GOA release. After just fetch-gene --force, 30 rows were newly
seeded and reviewed, and 18 rows retired by GOA were dropped (the validator treats
carrying them as an error). Among the additions are three experimental annotations the
review had no row for at all:

Most of the other 27 are additional per-interactor GO:0005515 rows, removed per
convention.

Other changes

Open questions

Recorded in suggested_questions. Two flagged here: