Biomolecular Condensates

Bottom line: membraneless compartments such as stress granules, P granules,
the nucleolus and the phagophore assembly site are heavily annotated as
locations, but GO and this repository almost never record which proteins build
them. This cross-cutting project audited that gap with a regenerable script:
238 gene folders carry a condensate-space CC term, yet only 9 carry
molecular condensate scaffold activity (GO:0140693), and reviewers had
already downgraded 146 of 399 (37%) reviewed condensate-space annotations
without a shared rule. It then re-reviewed every annotation in the corpus to
GO:0034045 phagophore assembly site membrane and its two related terms
(59 assertions across 21 gene folders, feeding GO issue #29437). Every
previously reviewed assertion had been accepted; after re-review, all 31
first-pass assertions across 14 genes moved to MODIFY (26) or
MARK_AS_OVER_ANNOTATED (5). GO has since obsoleted GO:0034045 and added
phagophore membrane (GO:7770114). The five working principles below are
still drafts, and the proposed calibration batch of scaffold genes has not
been run.

Membraneless compartments — stress granules, P granules, P-bodies, the nucleolus, PML
bodies, nuclear speckles, the phagophore assembly site — are among the most heavily
annotated structures in this corpus and among the least informatively annotated. This
project is the cross-cutting home for that problem: it does not own a condensate, it owns
the curation question that every condensate raises.

That question is simple to state. Being in a condensate is a location. Making one is a
function. GO annotation, and this repository along with it, records the first almost
exclusively and the second almost never.

The state of the corpus

A scan of every *-goa.tsv and *-ai-review.yaml in the repository (see the
GO and annotation audit, regenerable from a
committed script) gives the shape of the problem:

Eighty-nine genes are placed in the nucleolus; nine genes in the entire corpus are said
to scaffold any condensate at all. The asymmetry is not a curation backlog — it reflects
GO's own shape, where the compartments are richly subdivided and the activities that build
them are represented by a single molecular-function term.

Reviewers have already registered their discomfort without being asked to. Of 399 reviewed
condensate-space annotations, 146 (37%) were downgraded or worse — 121
KEEP_AS_NON_CORE, 17 MARK_AS_OVER_ANNOTATED, 8 REMOVE. Nucleolus alone accounts for
45 non-core, 9 over-annotated, and 3 removed. That is a consistent, repository-wide signal
that condensate localization, taken alone, is being judged uninformative about function —
gene by gene, with no shared framework behind it. Supplying that framework is what this
project is for.

Three ontology problems

1. GO has no class for "biomolecular condensate." The nearest candidate,
GO:0043228 membraneless organelle, cannot serve: the ribosome (GO:0005840) and the
cytoskeleton (GO:0005856) are both descendants of it. Neither is a phase-separated
condensate. So there is no term that picks out the set this project is about, and no way
to ask "which of my genes are condensate proteins?" without a hand-curated list — which is
exactly what the audit script has to maintain.

2. Condensates that are not classified as such. GO:0000407 phagophore assembly site
is not under membraneless organelle, though the PAS was shown in 2020 to be a liquid-like
Atg-protein condensate (PMID:32025038). The
inverse error to problem 1: a real condensate sitting outside the grouping. See
MODULE:phagophore_assembly_site, which models it as one.

3. Terms that assume a membrane the structure does not have. GO:0034045 phagophore
assembly site membrane defines "a cellular membrane associated with the phagophore assembly
site" — but the PAS is a condensate, and the term additionally carries phagophore and
isolation membrane as related synonyms, which are the name and synonym of the separate
term GO:0061908. Two cellular components, the same two strings. Recorded in full as a
knowledge gap on the PAS module.

Working principles

These are the commitments this project proposes; they are drafts until tested against a
batch.

1. Scaffold, client, or regulator — say which. A protein that phase-separates and
   nucleates the compartment (GO:0140693) is doing something categorically different from
   one that partitions into it. Reviews should decide which, and say so in
   core_functions, not leave a bare CC term to imply it.
2. Condensate CC alone is rarely a core function. The corpus already behaves this way
   (37% downgraded). Make it explicit rather than rediscovered per gene.
3. Prefer the material claim over the compartment claim. "Drives phase separation of X"
   is checkable; "localizes to structure Y" often reflects a fixation artefact or an
   overexpression phenotype.
4. Demand condensate-grade evidence. In-vitro droplet formation at non-physiological
   protein or salt concentrations, and puncta in overexpressing cells, are weaker than FRAP
   recovery, 1,6-hexanediol sensitivity with controls, and endogenous-tag imaging. Flag the
   distinction in review.reason.
5. Model condensates as modules, not just locations. A condensate has composition,
   assembly, and disassembly. MODULE:phagophore_assembly_site is the first worked example
   in this repository.

Existing assets

Relationship to the per-condensate projects

STRESS_GRANULES and CAEEL_P_GRANULES stay standalone; they are not folded in as sub-pages.
Each is a domain project with its own species scope, gene list, and disease framing, in the
same mould as PEROXISOME or ER_PHAGY, and CAEEL_P_GRANULES is already in progress with
reviewed genes — absorbing it would bury finished work and break its index entry. This page
is the cross-cutting layer above them: shared principles, shared ontology issues, and the
corpus-wide audit that no single condensate project would produce.

The P-granule work is also the methodological precedent worth generalising. It is the only
place in the corpus where curators asserted GO:0140693 as a new annotation rather than
accepting an existing one — meg-2, meg-3, meg-4, and pgl-2 all carry NEW, four of the five
NEW scaffold annotations in the repository. That is principle 1 already being applied,
before it was written down.

Proposed first batch

Genes that already carry GO:0140693 are the natural calibration set: small, cross-species,
and already reviewed, so the batch tests the principles rather than the pipeline.

• SQSTM1 — six scaffold annotations, five IDA, all ACCEPT; also the corpus's only
  GO:0140694 gene. The best-supported scaffold in the repository and the reference case.
• NFE2L2, LGALS3 — IDA scaffold annotations, ACCEPT; test whether the evidence
  meets principle 4.
• TP53 and its orthologues — the mouse and rat orthologues were reviewed
  KEEP_AS_NON_CORE and MARK_AS_OVER_ANNOTATED respectively, both on ISS/ISO. A live
  disagreement about whether a scaffold function propagates by similarity at all.
• Ccnt1, mid1 — non-human scaffold annotations; check taxon-appropriateness.
• TARDBP — heavily condensate-associated and disease-relevant, and a bridge to
  STRESS_GRANULES.

Open questions

• Should GO:0140693 have children distinguishing homotypic self-assembly from
  heterotypic client recruitment? Nine genes is too few to tell; the batch should decide.
• Is a "biomolecular condensate" CC grouping class worth requesting, given that the
  membrane/no-membrane axis already exists and cuts differently?
• What is the minimum evidence standard for a scaffold assertion, and should reviews record
  it structurally (e.g. as a knowledge-gap boundary) rather than in prose?
• Do IEA/ISS/ISO scaffold annotations survive scrutiny anywhere? Of the 22 in the corpus, 8
  are non-experimental and they account for every non-ACCEPT outcome.

Slides

• Slides (Marp source: CONDENSATES-slides.md) — AI generated