Readout → annotation rubric (ASSAY_TO_FUNCTION)

A curator-facing rubric for deciding what GO annotation a convergent
phenotypic readout licenses. Grounded in the publications-corpus mining (see
../ASSAY_TO_FUNCTION.md); machine-readable form in
rubric.yaml.

The core rule

A readout that reports the state of a process P (ROS level, caspase
activity, UPR activation, ΔΨm, Ca²⁺, pH, viability…) measures a downstream
consequence, not the gene product's molecular activity. So a positive readout
after perturbing gene G licenses at most:

• aspect BP (or CC for organelle-health probes) — never MF;
• framing "response to P" / "regulation of P" — not a direct effector role;
• default non-core.

Promote to core only when independent evidence (genetics, biochemistry,
structure) places G in the recognized machinery / sensor set for P — not
merely as a perturbation that moves the needle.

This is what curators are already doing in the corpus: the same readout yields
ACCEPT for the machinery and KEEP_AS_NON_CORE / MARK_AS_OVER_ANNOTATED
for genes acting indirectly.

Decision procedure

Is the evidence a state/phenotype readout (not a molecular assay of G)?
│
├─ No  → ordinary evidence rules apply (MF/direct annotation may be fine).
│
└─ Yes → 1. Reject any MF term sourced from this readout.
              (Exception: a transcriptional reporter for a bona fide
               DNA-binding TF — MF TF-activity is legitimate.)
         2. Use a "response to P" / "regulation of P" BP term at the
            altitude the readout DIRECTLY reports — not a more distal or
            tissue-specific child.
         3. Is G in the recognized core machinery / sensor set for P?
              ├─ Yes → core annotation OK (ACCEPT).
              └─ No  → KEEP_AS_NON_CORE (or MARK_AS_OVER_ANNOTATED if the
                       gene's real function clearly lies elsewhere).

Consolidated catalog (all 60 classes)

The catalog now spans 60 readout classes across six mining batches. The full
machine-readable summary (per-class proximity/convergence, aligned-annotation
GO-aspect counts, %MF, licensing) is auto-generated:

• reports/catalog_summary.tsv — one row per class
• reports/catalog_table.md — the same as a Markdown quick-reference (molecular vs
  phenotypic), regenerated from the catalog + mined matches
• reports/proximity_axis.png — the summary figure

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

The headline, in one line: across thematically-aligned annotations, the GO
aspect a readout licenses is set by proximity —

molecular readouts → 77% MF (567/738); phenotypic hubs → 8% MF (90/1087).

…and that 8% is almost entirely the legitimate TRANSCRIPTIONAL_REPORTER → DNA-binding-TF-activity exception (MF 72 of the 90); excluding it, phenotypic MF
≈ 2%. The one molecular outlier is PROTEASOME_ACTIVITY (2% MF) — its aligned
terms are the catabolic-process / complex, i.e. it reads as proteostasis
machinery rather than a bare endopeptidase MF.

The curated quick-reference below keeps the headline over-annotation hubs with
their machinery discriminators; see catalog_table.md for the exhaustive list.

Quick reference

* Ca²⁺-binding MF (EF-hand) can be justified by independent structural/binding
evidence, not by the imaging readout itself.

The pathway-reporter rows (Wnt/NF-κB/Notch/Hippo/hypoxia) are context rather
than non-core because, like a transcriptional reporter, the same readout
legitimately reports the core output of a bona fide pathway component (a
Frizzled receptor, RelA, a HIF subunit) — promote to core only for those, demote
for genes that merely perturb the reporter.

The molecular contrast: rubidium (⁸⁶Rb⁺) flux

Not every ion readout is a convergent hub. Rb⁺ flux (⁸⁶Rb⁺ efflux/uptake, the
classic K⁺-channel/transporter assay using Rb⁺ as a K⁺ congener) sits at the
molecular / low-convergence end of the proximity axis: it is a near-direct
measure of the gene product's own transport activity, so — unlike Ca²⁺ imaging (a
second-messenger hub that licenses at most a BP term) — it legitimately licenses
an MF channel-activity term (potassium channel activity, GO:0005267). It is
the positive-control mirror image of the phenotypic hubs.

Caveat (the same proximity logic in reverse): Rb⁺ flux is direct only for the
pore-forming channel. If flux moves because an upstream regulator, β-subunit,
or trafficking factor was perturbed, the inference is exactly as indirect as the
hubs — BP-only, default non-core.

Corpus note: ⁸⁶Rb is detected in 33 cached papers but is almost never the cited
original_reference_id of a reviewed annotation (only 1, and it is not a K⁺
channel), so the corpus is currently under-powered to demonstrate the MF licensing
empirically. This re-illustrates the first-pass finding that MF annotations cite
structural/biochemical references, not functional-flux assays.

Molecular MF-licensing controls (the positive-control set)

Where Rb⁺ flux was too niche, common molecular assays demonstrate the MF side of
the proximity axis directly. Aligned-annotation aspect (canonical join):
ChIP/EMSA → MF 136, in-vitro kinase assay → MF 106, GTPase/GAP/GEF →
MF 55, in-vitro ubiquitination/E3 → MF 39 (+CC 24 for the ligase complexes),
electrophysiology → MF 18. These are the molecular mirror of the hubs: a
readout of the gene product's own activity (DNA binding, phosphotransfer, GTP
hydrolysis, ubiquitin transfer, ion conduction) legitimately licenses an MF term.
Pairings to keep in mind when curating:

• ChIP/EMSA (MF, direct DNA binding) ↔ transcriptional reporter (BP, pathway output)
• electrophysiology (MF, channel activity) ↔ Ca²⁺/Rb⁺ imaging-flux (BP, ion state)
• in-vitro kinase assay (MF) ↔ phospho-Western of a downstream substrate (BP, pathway state)

Same caveat as Rb⁺ flux: these are direct for the assayed protein; if the readout
moves because an upstream regulator was perturbed, the inference is indirect.

Worked contrasts from the corpus

Each pair shows the same readout licensing a core annotation for the
machinery vs. an over-annotation for a gene acting indirectly:

• Apoptosis: BCL2 → negative regulation of apoptotic process (ACCEPT)
  vs Akt1 → negative regulation of intrinsic apoptotic signaling (NON_CORE).
• Proliferation: CDK1 → G2/M transition of mitotic cell cycle (ACCEPT)
  vs ACTB → positive regulation of cell population proliferation (OVER_ANNOTATED).
• ROS: TP53 → positive regulation of ROS metabolic process (ACCEPT)
  vs PEX10/PEX13 (peroxins) → cellular response to ROS (OVER_ANNOTATED).
• UPR: ATF4 → PERK-mediated UPR (ACCEPT) vs HSPA1A → ER UPR
  (NON_CORE); and IRE1 → unfolded protein binding (MF, OVER_ANNOTATED) —
  a textbook reporter-driven MF over-call.
• Autophagy: AMBRA1 → autophagosome assembly (ACCEPT) vs ABI2 →
  mitophagy (OVER_ANNOTATED).
• Transcription (the MF exception): ATF2/ASCL1 → DNA-binding
  transcription factor activity (ACCEPT — genuine TFs) vs AIP →
  transcription coactivator activity (OVER_ANNOTATED).

Extended set (second-pass readout classes):

• Cell migration / invasion: CLTC-style adhesion/cytoskeletal machinery and
  guidance ligands whose signature is motility stay core — CCL11/eotaxin →
  cell chemotaxis and PDGFA/PDGFB → cell migration / chemotaxis are
  ACCEPT (these are already in core_functions: a chemokine's defining output is
  chemotaxis). Contrast STAT3 → positive regulation of cell migration (a
  transcription factor; migration is a downstream transcriptional consequence,
  not in the motility machinery) — the genuine non-core/borderline case.
• DNA-damage foci: BRCA1/BRCA2/CHD1/RAD18 → double-strand break
  repair / DNA repair complex (ACCEPT — DDR machinery) vs a gene that merely
  shows γH2AX foci after perturbation (non-core "response to DNA damage").
• Endocytosis: CLTC (clathrin) / TFRC (transferrin receptor) →
  clathrin-dependent endocytosis / receptor-mediated endocytosis (ACCEPT —
  the machinery and the internalized receptor) vs a cargo that is merely taken
  up (non-core).
• Wnt reporter: CTNNB1/AXIN1/FZD7 → canonical Wnt signaling pathway
  (ACCEPT — pathway machinery) vs a gene that only shifts TOPFlash (non-core).
• Senescence / NF-κB / hypoxia: TP53 → cellular senescence, TRAF6 →
  NF-κB signal transduction, ARNT/HIF-1β → response to hypoxia are ACCEPT
  because the gene is the pathway's core effector/transducer — the reporter
  reports its own output.

Caveat: dedicated signaling ligands (signature vs incidental)

The "phenotypic hub readout ⇒ non-core" rule, and the flagger's
indirect_ligand discriminator, over-fire on dedicated cytokines and growth
factors. For a gene whose entire biological purpose is to regulate a process
(VEGFA→endothelial proliferation; IL21→B-cell/Tfh responses), the regulated
process is core even though it is mechanistically downstream of receptor
signaling. "Cytokine activity" alone is nearly contentless, so the regulated
processes are the informative, identity-defining annotations.

The discriminating axis is therefore signature vs incidental, not
ligand-vs-not:

• Signature — the process the gene is known/named for, and that loss-of-
  function abolishes → keep core (VEGFA endothelial proliferation; IL21 germinal
  center / Tfh / B-cell responses).
• Incidental / generic / tissue-specific-secondary / context-dependent →
  demote (HMGB1→CXCR4 calcium; VEGFA→generic anti-apoptosis; IL21→T-cell
  proliferation is borderline and was deferred to expert review, issue #1418).

Mechanistic directness still matters for the ligand vs receptor distinction
(knockout necessity ≠ the ligand directly performing the process), but it
must not be applied so bluntly that a dedicated cytokine's signature outputs are
stripped to non-core.

How this was derived

Across 722 thematically-aligned (annotation, source-paper-readout) pairs, the
GO aspect of hub-readout annotations is overwhelmingly BP/CC with ~zero MF
(transcriptional reporters excepted, and there the MF is legitimate). The
distinguishing action is non-core demotion, not removal — see the action
and aspect tables in ../ASSAY_TO_FUNCTION.md. The
"core only if in the machinery" discriminator is read directly off the
ACCEPT-vs-downgrade contrasts above.

Extension validates the constraint. Adding 12 new readout classes (migration/
invasion, adhesion, endocytosis, secretion, metabolic flux, DNA-damage foci,
senescence, and pathway reporters for Wnt/NF-κB/Notch/Hippo/hypoxia) reproduced
the same pattern: every new aligned 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 each shows elevated non-core demotion (migration 16/35,
membrane-trafficking 7/14, Wnt 8/17). Re-review of the standing-ACCEPT
candidates again found the flagger's precision-on-accepted-calls low: they are
either the machinery (BRCA1/2, CLTC, TFRC, CTNNB1, TRAF6, ARNT, TP53 —
correctly core) or a signature output of a dedicated ligand (CCL11/PDGF
chemotaxis — correctly core, already in core_functions). The machinery and
signature-vs-incidental discriminators are doing the real work.