ASSAY_TO_FUNCTION

Which experimental readouts reliably support gene-function annotation, and which drive over-annotation?

Motivation

Functional genomics papers infer gene function through a chain:

perturb gene G → readout R moves → conclude G is "involved in" process P

The reliability of that final R → P step varies enormously, but GO evidence
codes (IDA, IMP, IGI, …) do not capture it. An IMP read out with a 5×UPRE
reporter and an IMP read out with a reconstituted enzyme assay carry the same
evidence code, yet one is a distal, convergent phenotype and the other is a
direct measurement of molecular activity. This project tries to make that
hidden axis explicit and to measure, against the curated corpus, which readout
classes are associated with annotations that curators later downgrade.

Conceptual framework

Two independent axes govern R → P reliability:

1. Proximity — does the readout measure the gene product's own molecular
   activity (molecular) or a downstream cellular consequence (phenotypic)?
2. Convergence — is the readout a fairly specific signature of process P
   (low), or a hub that many upstream inputs feed into (high)?

Over-annotation risk is highest in the phenotypic + high-convergence
quadrant. Classic convergent hubs: ROS, caspase activation, the UPR,
intracellular Ca²⁺, pH, mitochondrial membrane potential. A gene that moves one
of these reporters could be a direct effector or anything that perturbs the
process indirectly — so a positive readout licenses, at best, a cautious
response to / regulation of BP term (often non-core), and should never
by itself drive an MF or core-function call.

Example readouts (the seed cases that motivated this)

The full, editable catalog lives in
ASSAY_TO_FUNCTION/readout_catalog.yaml.

Method (first pass — mining the curated reviews)

ASSAY_TO_FUNCTION/mine_readouts.py walks all genes/**/*-ai-review.yaml
files. For each annotation it builds an evidence text from the curator's
review.summary, review.reason, and supported_by[].supporting_text, then
matches it against the regex patterns in the catalog. Each match becomes a row
joining the readout class to the reviewer's action.

Reliability is summarised two ways, over reviewed annotations only
(UNREVIEWED / UNDECIDED / PENDING excluded from the denominator):

• pct_removed_or_overann — hard signal: REMOVE + MARK_AS_OVER_ANNOTATED
• pct_any_downgrade — soft signal: also includes KEEP_AS_NON_CORE + MODIFY

uv run python projects/ASSAY_TO_FUNCTION/mine_readouts.py

Outputs land in ASSAY_TO_FUNCTION/reports/:
readout_matches.tsv (one row per match), readout_action_crosstab.tsv
(the table below), and matched_string_counts.tsv (QC).

First-pass results

Scanned 1,971 review files / 75,931 annotations; 1,736 readout matches.

Interpretation (preliminary — under-powered)

• Directional support, not proof. The two phenotypic-hub classes with
  non-trivial N — apoptosis/caspase (49% any-downgrade) and mitochondrial
  membrane potential (42%) — sit well above the molecular controls
  (in-vitro enzyme 24%, direct binding 18%). But UPR, autophagy, and
  transcriptional reporters are no higher than the molecular baseline, and
  several classes have N < 10. No strong claim is warranted yet.
• The review prose is the wrong corpus for assay names. Curators write
  synthesis ("phosphorylates ACC", "crystal structure shows…"), not methods.
  Probe brand names are almost absent: CellROX/DCFDA/MitoSOX matched zero
  review texts. The molecular classes dominate by volume only because "crystal
  structure", "cryo-EM", and "in vitro" are phrases curators naturally use.
• Implication: to actually measure how assays are interpreted, we must mine the
  publications (methods sections), not the review summaries — see next steps.

Methodology lessons (substring bugs caught in QC)

Keyword mining is treacherous; the QC dump caught three false-positive sources
that would have completely inverted the headline numbers:

• HyPer (H₂O₂ biosensor) matched "hyper-activation", "hyper-expressed"
  → inflated oxidative stress by ~365. Fixed with case-sensitive (?-i:HyPer).
• ERSE (ER stress element) matched "diverse", "reverse", "adverse"
  → inflated UPR from ~29 to 609. Fixed with a left word boundary.
• MTS (viability dye) matched "MTs" (microtubules) and the
  mitochondrial-targeting-sequence abbreviation → fixed by requiring
  "MTS assay/reagent/reduction".

Always inspect matched_string_counts.tsv before trusting a class total.

Second pass — mining the publications corpus

ASSAY_TO_FUNCTION/mine_papers.py does what the first pass showed was needed:
the unit of analysis is now (annotation, readout-class-used-in-the-source-paper).
For each PMID-backed annotation it resolves the cached publications/PMID_*.md,
detects which readout classes that paper uses, attaches the GO aspect
(MF/BP/CC, from the GOA TSVs), and flags thematic alignment — whether the
annotation's GO term is actually about the process the readout reports
(GO id in commonly_overmapped_to, or label matches aligned_label_regex).

uv run python projects/ASSAY_TO_FUNCTION/mine_papers.py

Coverage: 75,931 annotations; 36,660 PMID-backed; 36,449 papers resolved
(129 PMIDs not cached). 16,682 paper-level readout matches; 722 thematically
aligned (the high-precision subset where the readout plausibly drove the
annotation).

Performance note: detection is a two-stage filter — a cheap literal substring
screen (SCREEN in the script, a necessary superset per class) gates the
expensive IGNORECASE regex, which runs only on the rare screen hits. Screens
were validated to lose zero matches vs. regex-only on a 350-paper sample
(including the 50 largest). Detection is memoized per unique PMID.

Headline result: the aspect constraint holds

For thematically aligned annotations, what GO aspect does each hub readout
license?

This directly supports the core hypothesis: phenotypic-hub readouts almost
never license MF annotations. Caspase, ROS, autophagy, viability, and UPR
readouts produce BP terms (ΔΨm probes produce mostly CC = mitochondrion
localization). The one exception is transcriptional reporters (MF 71) — but
inspecting those rows, they are bona fide TF-activity terms (DNA-binding transcription factor activity, transcription cis-regulatory region binding)
on genuine transcription factors, mostly ACCEPTed. A luciferase reporter
legitimately supports MF for a TF; it is not over-annotation there.

The real over-annotation mode: correct-but-non-core, not wrong

Action breakdown for aligned annotations (the strong-link subset):

The key correction to the first-pass intuition: hub readouts are not removed
more often than molecular evidence (hard rm/OA% is comparable — molecular
controls run ~27% in the weak-link view). Instead their characteristic failure
mode is demotion to non-core: viability/proliferation (58/90 → NON_CORE)
and apoptosis (34/63) annotations are rarely wrong, but are overwhelmingly
judged peripheral. That is the precise, defensible sense in which these
readouts "over-annotate": they inflate the annotation set with correct-but-
peripheral process terms rather than producing outright errors.

Mitochondrial-membrane-potential readouts are the exception with a genuinely
elevated hard downgrade rate (21%), often via MODIFY (31/165) — ΔΨm is a
non-specific organelle-health readout that gets refined to more proximal terms.

Caveats

• Weak link. A paper using assay R does not prove the specific annotation
  was based on R. Thematic alignment mitigates this (722 high-precision rows)
  but cannot fully establish causation.
• The any-downgrade rate is high (40-60%) for every class including
  molecular controls, because it is dominated by KEEP_AS_NON_CORE/MODIFY,
  which are refinements, not errors. The hard rm/OA% and the aspect/non-core
  breakdowns are the discriminating signals, not the raw downgrade rate.

Consolidated catalog (deliverable)

ASSAY_TO_FUNCTION/consolidate.py folds the 60-class catalog and the mined
aligned matches into a single summary — reports/catalog_summary.tsv,
reports/catalog_table.md (a complete molecular-vs-phenotypic quick-reference),
and reports/proximity_axis.png (the summary figure).

uv run --with matplotlib python projects/ASSAY_TO_FUNCTION/consolidate.py

One-line result: across thematically-aligned annotations, the licensed GO
aspect tracks proximity — molecular readouts 77% MF (567/738) vs phenotypic
hubs 8% MF (90/1087), a ~10× separation over 60 assay families. The phenotypic
8% is almost entirely the legitimate TRANSCRIPTIONAL_REPORTER → TF-activity
exception (72/90; ≈2% without it); the lone molecular outlier is
PROTEASOME_ACTIVITY (2% MF — read as proteostasis process/complex). See
RUBRIC.md for
the figure and table.

The rubric (deliverable)

The grounded, curator-facing rubric is in ASSAY_TO_FUNCTION/RUBRIC.md
(narrative + decision procedure + worked corpus contrasts) with a
machine-readable companion in ASSAY_TO_FUNCTION/rubric.yaml.

Core rule: a convergent phenotypic readout licenses at most a BP (or CC)
"response to / regulation of P" term, never MF, defaulting to non-core —
promote to core only when independent evidence places the gene in P's recognized
machinery. The single MF exception is a transcriptional reporter for a bona fide
DNA-binding TF.

The flagger (operationalized rubric)

ASSAY_TO_FUNCTION/flag_candidates.py
applies the rubric to the mined matches and emits a prioritized re-review
worklist (reports/flagged_candidates.tsv). It flags only annotations that are
still standing (not already downgraded) and thematically aligned to a hub
readout, so precision stays high. Two tiers:

• Tier 1 — MF from a hub readout (a state readout cannot license molecular
  function). The legitimate transcriptional-reporter→TF-activity case is
  excluded.
• Tier 2 — a core (ACCEPT) hub-aligned BP/CC call, where the rubric
  default is non-core; verify the gene is in the recognized machinery.

uv run python projects/ASSAY_TO_FUNCTION/flag_candidates.py

Current run: 296 candidates (5 Tier 1, 291 Tier 2) (post Tier-1 binding-MF calibration — see Tier-1 re-review outcome below for the 7→5 reduction). The Tier-1 set is
precisely the reporter-driven over-annotation pattern the rubric predicts —
transcription coactivator/corepressor activity claimed from luciferase
reporters for coregulators that are not sequence-specific TFs (CTNNB1/
β-catenin, NOTCH1, SIRT1, HMGB1), plus Ca²⁺-binding MF terms (Calm2, HRC) that
should rest on EF-hand/structural evidence rather than Ca²⁺ imaging. Tier 2 is a
larger triage queue dominated by MITO_MEMBRANE_POTENTIAL (85, mostly generic
mitochondrion), TRANSCRIPTIONAL_REPORTER (72), and AUTOPHAGY_FLUX (70).

These are re-review candidates, not asserted errors.

Tier-1 re-review outcome (loop closed)

All 7 original Tier-1 candidates were re-reviewed against their actual evidence
(see ASSAY_TO_FUNCTION/REREVIEW_TIER1.md).
Outcome: none is a clean readout-driven over-annotation — six KEEP, one
(SIRT1 corepressor activity) a soft non-core suggestion. Three calibration
findings fed back:

1. Binding MF is a direct activity, not a readout consequence — Calm2/HRC
   ("calcium … binding") were false positives. The flagger now excludes
   *binding MF from Tier 1 (7 → 5), and the rubric's "never MF" rule is scoped
   to regulatory/process-like MF, not binding/catalytic MF.
2. Coregulator MF is legitimate for genuine coregulators (β-catenin, NICD);
   the discriminator vs. the corpus's true over-annotation (AIP → coactivator)
   is machinery membership, not aspect.
3. The standing-only filter works: the real over-annotation (AIP) was
   already MARK_AS_OVER_ANNOTATED, so was correctly not re-flagged. The
   flagger's marginal value is therefore highest on unreviewed annotations and
   the Tier-2 queue, not on re-litigating accepted MF calls.

Tier-2 triage + the machinery discriminator

The Tier-2 queue (core ACCEPT on a hub-aligned BP/CC term) is large and, on
its own, low precision: re-reviewing the top VIABILITY_PROLIFERATION class
showed the queue mixes genuine machinery (CDK1, MYC, RB1, TP53 — correctly core)
with indirect cases (IL21, PDGFB, VEGFA). See
ASSAY_TO_FUNCTION/REREVIEW_TIER2.md.

The flagger now (a) ranks Tier 2 by each readout class's empirical any-downgrade
rate, and (b) adds a machinery discriminator: a candidate is tagged
indirect_ligand when the gene's own MF is a secreted signaling ligand
(cytokine/growth factor/hormone/chemokine), because then any cellular process it
drives is downstream of receptor signaling — the strongest computable non-core
signal. This cleanly isolates 6 high-precision non-core candidates (IL21, PDGFB,
VEGFA×2, HMGB1×2) at the top of the queue while the cell-cycle machinery sinks.

uv run python projects/ASSAY_TO_FUNCTION/flag_candidates.py --target accepted

Annotation edits made (curation output)

The analysis was carried through to actual YAML edits — see
ASSAY_TO_FUNCTION/EDITS.md for the full table
with gene descriptions and rationale. Summary: 4 genes / 6 annotation records
downgraded ACCEPT → KEEP_AS_NON_CORE where a process is genuinely
downstream of the gene's core MF:

• PDGFB — glomerular mesangial cell proliferation (GO:0072126) — downstream of PDGFR signaling
• HMGB1 — cytosolic calcium (GO:0007204) — downstream CXCR4 signaling
• Sirt2 — positive regulation of cell division (GO:0051781) — downstream of deacetylase activity
• VEGFA — negative regulation of apoptotic process (GO:0043066) — downstream Akt survival signaling

Deliberately kept as-is (flag was a false positive): VEGFA endothelial-cell
proliferation (defining function), SIRT1 corepressor activity, the
binding/coactivator MF terms from the Tier-1 re-review, and PDGFB mesangial
proliferation (a reversal was considered but rejected — knockout necessity is
not mechanistic evidence of direct proliferation regulation). All edited files
re-validate cleanly.

Deferred to expert review (UNDECIDED): IL21 positive regulation of T cell proliferation (GO:0042102) is genuinely borderline — a real but weak,
context-dependent effect versus IL21's signature B-cell/Tfh axis. Rather than
guess, the two IDA annotations were set to UNDECIDED (paper-backed) and the
term removed from core_functions pending
issue #1418.
This exposed a needed rubric refinement: the "signaling-ligand ⇒ indirect"
discriminator over-fires on dedicated cytokines/growth factors, whose
regulated processes can be core. The better axis is signature vs incidental
(see RUBRIC.md).

Catalog extension — 12 new readout classes

The catalog was extended beyond the seed hubs to cover more phenotypic assay
families: cell migration/invasion (scratch, Transwell, Boyden, Matrigel),
cell adhesion/spreading, membrane trafficking/endocytosis (transferrin/
FM4-64/dextran uptake), secretion/degranulation (LDH release, CD107a,
β-hexosaminidase), metabolic flux (Seahorse/OCR/ECAR, 2-NBDG glucose uptake),
DNA-damage foci (γH2AX, comet, 53BP1/RAD51 foci), senescence (SA-β-gal,
SASP), and pathway-specific transcriptional reporters for Wnt (TOPFlash),
NF-κB, hypoxia (HRE/HIF), Notch (RBP-J/CSL), and Hippo (TEAD/GTIIC).
Each class got an aligned_label_regex, commonly_overmapped_to GO IDs, regex
patterns, and a necessary-superset literal SCREEN entry.

mine_papers.py now also emits reports/paper_matched_string_counts.tsv — a
corpus-level matched-substring QC (counted once per paper), the publications
analogue of the prose miner's QC and the place substring bugs actually surface.

QC caught one substring bug

\bOCR\b (oxygen consumption rate) matched the C. elegans ocr-2 TRPV-channel
gene (19 false hits). Dropped bare OCR; the class now relies on the spelled-out
"oxygen consumption rate" / Seahorse / ECAR. After the fix
paper_matched_string_counts.tsv is clean across all 12 new classes.

The aspect constraint generalizes

Aligned annotations per new class (publications corpus):

Every new class is BP/CC-dominant with ~zero MF (the sole MF, LRRK2 β-catenin
destruction complex binding, is a binding MF already non-core), and the
well-powered ones show the predicted elevated non-core demotion. The same
BP-not-MF, default-non-core regime holds.

Re-review: machinery vs signature, again

The standing-ACCEPT candidates in the new classes are, on re-review, almost all
correctly curated — the flagger's precision on accepted calls remains low:

• Machinery → core (KEEP): BRCA1/2, CHD1, RAD18 (DNA-damage); CLTC, TFRC
  (endocytosis); CTNNB1, AXIN1, FZD7 (Wnt); TRAF6 (NF-κB); ARNT/HIF-1β (hypoxia);
  TP53 (senescence) — the "core only if in the machinery" discriminator.
• Signature ligand outputs → core (DECLINE): the CELL_MIGRATION_INVASION
indirect_ligand cluster (CCL11, PDGFA/B, VEGFA, HMGB1) is exactly the
  signature-vs-incidental over-firing the rubric's caveat predicts — chemotaxis
  is a chemokine's job; PDGF migration is already in core_functions.
• Genuine non-core/borderline → defer: STAT3 → positive regulation of cell
  migration (GO:0030335, ×2 IMP) — a transcription factor whose migration role
  is a downstream transcriptional consequence, not motility machinery. High-profile
  and contested (oncogenic EMT driver vs downstream), so set ACCEPT → UNDECIDED,
  opened issue #1422,
  and staged + submitted a cited OpenScientist hypothesis job (see below) rather
  than unilaterally demoting. See ASSAY_TO_FUNCTION/EDITS.md.

A molecular positive-control class: rubidium (⁸⁶Rb⁺) flux

To probe the other end of the proximity axis, a RUBIDIUM_FLUX class was added
as molecular / low-convergence — the classic ⁸⁶Rb⁺ K⁺-channel/transporter
assay (Rb⁺ as a K⁺ congener). Unlike the Ca²⁺ imaging hub, Rb⁺ flux is a near-direct
measure of the channel's own activity and should license an MF channel-activity
term — the mirror image of the phenotypic hubs. QC is clean (the ⁸⁶Rb notation
matched 33 papers with no RB1-gene/rubidium-salt false positives — bare Rb is
deliberately excluded from the screen).

Honest result: under-powered to null (robust). Of the 33 ⁸⁶Rb papers, only one
is cited by any reviewed annotation (mTOR transmembrane transporter binding, not
a K⁺ channel — correctly not aligned), so there are zero aligned Rb-flux
annotations — and this holds even under the broader supported_by join below.
The 33 ⁸⁶Rb papers are essentially disconnected from the curated annotation
references. The class is kept as a correctly-implemented control that future
ion-channel-gene coverage would populate; the null re-confirms the first-pass
lesson that MF annotations cite structural/biochemical references rather than
functional-flux assays. Caveat for when it does populate: Rb⁺ flux is direct only
for the pore-forming channel — flux moved by perturbing a regulator/subunit is the
same indirect inference as the hubs.

Broader join: supported_by references (not just the primary)

mine_papers.py --include-supporting joins each annotation to readout usage
across all its cited papers (supported_by / additional_reference_ids), not
only original_reference_id, recording a ref_role (primary/supporting) per row.
Written to reports/with_supporting/ so the canonical strong-link analysis stays
intact.

Effect: PMID-backed annotations 37.6k → 47.2k, paper-readout matches 18.9k →
28.1k, thematically aligned 863 → 1,200 (+39%; 337 of the 1,200 from
supporting refs). Crucially the headline pattern strengthens — the phenotypic
hubs remain BP/CC-dominant with ~zero MF on the larger sample (apoptosis BP93/MF0,
autophagy BP186/CC30/MF0, DNA-damage BP58/CC7/MF0; TRANSCRIPTIONAL_REPORTER's MF
is the legitimate-TF exception). QC of the broader run is clean. The broader join
did not rescue RUBIDIUM_FLUX, making that null robust rather than an artifact
of the primary-only join.

The proximity axis, demonstrated both ways (molecular vs phenotypic)

The first extension showed phenotypic hubs license BP/CC and ~never MF. A second
batch of 10 classes was added to test the other prediction directly — that
molecular assays of the gene product's own activity license MF — by
including common, well-cited molecular readouts (not just the niche Rb⁺ flux):
electrophysiology, in-vitro kinase assays, GTPase/GAP/GEF assays, in-vitro
ubiquitination/E3-ligase assays, and ChIP/EMSA; plus five more phenotypic hubs
(differentiation, angiogenesis/tube-formation, phagocytosis, cell-cycle/flow,
barrier/TEER).

The aspect of thematically-aligned annotations splits exactly on the axis
(canonical join; QC clean — e.g. ChIP required ChIP+suffix so no potato/
microarray "chip" leaked, and the broad kinase screen still gated a precise
kinase assay regex):

This is the framework's central claim shown as a single within-corpus contrast:
a readout of the gene product's own activity (DNA binding, phosphotransfer,
GTP hydrolysis, ubiquitin transfer, ion conduction) licenses an MF term,
whereas a downstream phenotype (differentiation, a tube, an engulfed
particle, a cell-cycle profile, a resistance drop) licenses BP/CC and ~never
MF. Electrophysiology (MF 18) supplies the channel-activity positive control
that the niche RUBIDIUM_FLUX class could not. The split is even sharper under the
broader --include-supporting join (ChIP MF 212, kinase MF 153, GTPase MF 93,
ubiquitination MF 51, electrophysiology MF 35).

Curation: the new phenotypic-hub flags were all machinery or signature
on re-review — VEGFA→angiogenesis (signature), GATA3/SOX9 (master differentiation
TFs), RB1/BRCA1 (cell-cycle machinery) — so no edits were warranted (the
"core only if in the machinery / signature output" discriminators at work again).

Further coverage: proteostasis, lipid, redox, nucleic-acid handling

A fourth batch (8 classes, catalog now 43 readout classes total) extended into
proteostasis, lipid, redox, and nucleic-acid assays and reproduced the axis once
more. Molecular catalytic readouts license MF — NUCLEASE_ACTIVITY MF 20,
PROTEASE_ACTIVITY MF 15, LIPID_TRANSFER_FLIPPASE MF 5 (canonical; NUCLEASE MF 42 /
PROTEASE MF 33 with supporting refs). Phenotypic state readouts license BP —
PROTEIN_TURNOVER (CHX chase/half-life) BP 29, TRANSLATION_ASSAY BP 13,
LIPID_PEROXIDATION BP 12, REDOX_BALANCE BP 1 (under-powered). PROTEASOME_ACTIVITY
aligns BP/CC (BP 30, CC 12) — its label regex maps to the catabolic-process and
complex terms, i.e. it reads proteostasis function/location rather than a bare
endopeptidase MF.

QC clean, with three substring traps deliberately dodged: bare MDA (collides
with the MDA-MB cell-line series) → malondialdehyde; bare puromycin (a
selection antibiotic) → puromycin incorporation/SUnSET; cyclic-nucleotide
deferred to avoid cAMP→"hippocampus"/"campaign". The 27 new phenotypic-hub
flags are again all machinery (PSMA1/PSMB5 proteasome subunits, CUL3 cullin
E3, CDC37/PEX19 chaperones) — correctly core, no edits.

Fifth batch: epigenetic enzymes, immune readouts, second messengers

A fifth batch (9 classes; catalog now 52 readout classes) added epigenetic
writers/erasers, immune assays, and the second-messenger class — done safely —
and reproduced the axis a fifth time. Molecular catalytic readouts → MF:
ACETYLTRANSFERASE_DEACETYLASE MF 59, POLYMERASE_ACTIVITY MF 15, PHOSPHATASE MF 8,
METHYLTRANSFERASE MF 7, HELICASE MF 7 (with supporting refs: acetyl MF 76,
polymerase MF 26). Phenotypic state readouts → BP: HISTONE_MARK (H3K4me3 etc.)
BP 13, CYCLIC_NUCLEOTIDE_SIGNALING BP 10, CYTOKINE_PRODUCTION BP 7,
CYTOTOXICITY_KILLING BP 3.

The histone-mark vs methyltransferase/acetyltransferase pair is a nice epigenetic
echo of the reporter-vs-ChIP one: the enzyme assay (HAT/HDAC/HMT) licenses an
MF, while the mark state (H3K4me3, H4K16ac) is a downstream BP chromatin
readout. The cyclic-nucleotide class was added without the cAMP→"hippocampus"
substring trap by requiring an explicit assay/level/sensor context.

QC clean. The 11 new phenotypic-hub flags are again machinery or signature —
chromatin enzymes (RTT109 HAT, SET1 methyltransferase, CHD1 remodeler, ASF1
chaperone) and GPCRs whose cyclase-activating signaling is their signature
(ADRB2, Drd1) — correctly core, no edits.

The proximity axis has now held across five independent batches spanning ~30
molecular-vs-phenotypic assay families (43→52 classes).

Sixth batch: RNA-binding, transport, chaperone, glyco; EMT/stemness/aggregation/inflammasome

A sixth batch (8 classes; catalog now 60 readout classes) reproduced the axis
a sixth time. Molecular → MF: CHAPERONE_REFOLDING MF 49, GLYCOSYLTRANSFERASE
MF 12, RNA_BINDING_CLIP MF 8, TRANSPORTER_UPTAKE MF 7 (supporting refs: chaperone
MF 85, transporter MF 25, glyco MF 21, RNA-binding MF 14). Phenotypic → BP:
PROTEIN_AGGREGATION BP 22 (+MF 12 from amyloid-binding terms), INFLAMMASOME_
PYROPTOSIS BP 15, EMT_MARKERS BP 9, STEMNESS_SPHERE BP 7. RNA_BINDING_CLIP is the
RNA counterpart of CHROMATIN_CHIP (CLIP/RIP/REMSA → RNA-binding MF). QC clean; the
bare-RIP trap (receptor-interacting protein) was dodged by requiring RIP-seq /
RIP assay / RNA immunoprecipitation.

The 23 new phenotypic-hub flags are again machinery or signature: anti-aggregation
chaperones (BAG3, CLU, CRYAB — negative regulation of amyloid fibril formation
is their function), APP (the amyloid precursor itself), FZD7 (Wnt receptor →
stemness), and anti-inflammatory cytokines whose signature is inflammation
regulation (IL10, IL36RN) — correctly core, no edits.

Six batches, ~36 assay families, 12→60 classes: the proximity axis is robust.
Across every batch, a readout of the gene product's own activity (catalysis,
binding, transport) licenses MF, while a downstream phenotype licenses BP/CC and
~never MF; and every standing-ACCEPT flag re-reviewed to machinery or a signature
output, so the flagger's curation value remains on unreviewed annotations rather
than re-litigating accepted core calls.

Cited-adjudication complement: staged OpenScientist jobs

For borderline / signature-vs-incidental disputes, the project stages
openscientist.io hypothesis jobs as a cited literature complement (the pattern
from the IL21 #1418 run). A human-gated generator turns selected
flagged_candidates.tsv rows into committed, frontmatter-free prompt.md files:

# stage prompts (writes prompt.md, prints submit commands, NEVER submits)
uv run python projects/ASSAY_TO_FUNCTION/stage_hypotheses.py --discriminator indirect_ligand --max 5
uv run python projects/ASSAY_TO_FUNCTION/stage_hypotheses.py --gene STAT3 --go-id GO:0030335
# or: just assay-stage-hypotheses --gene STAT3 --go-id GO:0030335

It reuses scripts/gene_hypothesis_deep_research.py so staged prompts are
identical to what that tool submits, specialized to the "core function vs
downstream/context-dependent consequence?" question. A human reviews the staged
prompts and submits only the few worth a paid (~15–30 min) job. The STAT3 job was
run as the worked example; its verdict is posted to #1422 and treated as
hypothesis-generating synthesis to verify against the cited PMIDs — the annotation
stays UNDECIDED until an expert decides.

Next steps

1. Curator triage of the ranked flagged_candidates.tsv, starting with the
   indirect_ligand subset (highest-precision non-core candidates).
2. Expand the catalog with more probe vocabularies and the convergent
   process-term GO IDs each readout gets over-mapped to.
3. Generalize the discriminator beyond signaling ligands (e.g. transporters,
   structural proteins) — currently those indirect cases still need human
   judgement via the ranked queue.

Relationship to existing projects

• OVER_ANNOTATION_PATTERNS.md — pattern #4 ("indirect downstream process
  annotations") is the conceptual cousin; this project quantifies the
  assay-specific version of it.
• BIOSENSORS.md — unrelated (plant synthetic-biology biosensors).