Behaviour Annotation Project

MATURE PIPELINEFLAGSHIP

Species: mouse, human, rat, worm, yeast, DANRE, DROME, DAPPU

Genes: App STAT3 nphp-1 Casp3 Drd1 CRY lov-1 pkd-2 GCG daf-2 trpm7 Tuba1a Agtr1a Mtor Fyn

Behaviour Annotation Project

Bottom line: when a knockout changes how an animal behaves, the gene
gets annotated to behavior (GO:0007610) or a child term, even when its
molecular function sits many causal steps upstream. We mined every GOA file
and every review in the corpus for behaviour terms, wrote a four-step rubric
(remove if contradicted, accept if the gene acts in the circuit, keep as
non-core if distal, mark over-annotated if uselessly broad), spot-checked
every accepted row, and mapped 16 standardized behavioural assays (15 IMPReSS
types plus the Morris water maze) to the GO terms they can support. We did this because behaviour is the most distal and
most convergent readout there is, so it is the cleanest test of how reviewers
handle phenotype-driven annotations. A re-run of the miner in September 2026
finds 216 behaviour annotations in reviews, of which 197 were adjudicated as
core or not: 169 (86%) were downgraded and 28 accepted, mostly sensory
channels and receptors acting in the relevant neurons (worm lov-1, pkd-2,
tax-4), plus the fly clock protein CRY and the satiety hormone GCG. reports/REPORT.md
is regenerated from that re-run; the ~81% → 87% figures in the spot-check
section below are the June snapshot (146 adjudicated) that motivated it, and
the conclusion is unchanged.

Motivation

The Gene Ontology behavior branch (GO:0007610) sits at the extreme
organismal end of the biological-process hierarchy. A behaviour is an
integrated, whole-animal output — locomotion, feeding, mating, grooming,
circadian rhythm, a fear response. Almost any perturbation that reaches the
nervous system, or that compromises development, metabolism, or basic
cell biology, can shift one of these readouts.

The annotation chain that produces a behaviour term is the same convergent,
distal pattern flagged in ASSAY_TO_FUNCTION and
OVER_ANNOTATION_PATTERNS:

perturb gene G → animal behaves differently → annotate G to behaviour B

The evidence codes that dominate behaviour annotations (IMP from a mutant,
IGI from a genetic interaction) record that the phenotype is real and
reproducible — but they say nothing about how proximal the gene is to the
behaviour. A tubulin, a lysosomal peptidase, an angiotensin receptor, and a
ciliary scaffold can all earn a locomotory behavior annotation from a
phenotype assay, yet none of them is a "behaviour gene" in any mechanistic
sense. The behaviour is a downstream readout of a much more specific molecular
defect.

So the goal here is not to delete behaviour annotations wholesale. A
behaviour phenotype is legitimate evidence, and for a handful of genes
(neurotransmitter receptors, neuropeptides, circadian clock components) a
behaviour term genuinely is close to the core function. The goal is to
separate the core from the consequence: keep well-supported behaviour
annotations as non-core context, flag the distal ones as
over-annotations, and reserve REMOVE for the cases that are actually
contradicted (wrong paralog, wrong gene) rather than merely distal.

What the corpus already shows

Mined with BEHAVIOR/mine_behavior.py over every
*-goa.tsv (source annotations) and every *-ai-review.yaml (reviewer
decisions). The full tables regenerate into
BEHAVIOR/reports/REPORT.md.

Source surface. Behaviour terms in the corpus GOA files are mostly
phenotype-driven: IMP + IGI account for just over half (114 of 209, 55%)
of behaviour annotations; most of the rest is electronic or inferred by
similarity (IEA 39, ISS 26, ISO 14), and only 3 are IDA (direct assay). The
most common term is the broad locomotory behavior (GO:0007626, 26 rows) by a
wide margin, followed by a cluster of terms at 7–11 rows each: mating behavior, social behavior, behavioral response to pain, drinking behavior, chemosensory behavior, adult locomotory behavior,
thermosensory behavior, and circadian behavior (counts from the September
2026 re-run of the miner).

Reviewer decisions. Of the behaviour annotations reviewers have adjudicated
as core-vs-not (197 of 216 in the September 2026 re-run, excluding the 10 NEW
proposed terms, which add rather than downgrade, and 9 UNDECIDED), 169 (86%)
were downgraded — kept as non-core, marked as over-annotated, or removed — and
only 28 were ACCEPTed as a core function:

Action Meaning for a behaviour term Share
KEEP_AS_NON_CORE Real phenotype, distal to molecular function dominant (147 of 216, ~68%)
ACCEPT Behaviour genuinely near the core (e.g. receptors, clock genes) minority
MARK_AS_OVER_ANNOTATED Too broad / too distal to be useful small
REMOVE Contradicted — wrong gene/paralog or not supported small

This is exactly the signature of a benign-but-pervasive over-annotation
pattern: the annotations are mostly not wrong, but they are mostly not core.

Exemplars from completed reviews

These are real decisions already in the corpus — concrete illustrations of the
"keep as non-core, it's a downstream readout" call:

The contrast between Tuba1a/tpp1 (distal → non-core) and Agtr1a (wrong paralog →
remove) is the core curation distinction this project sharpens.

Curation guidance (working rubric)

For a behavior (GO:0007610-descendant) annotation, ask:

  1. Is it contradicted? Wrong gene, wrong paralog, or the cited evidence
    actually attributes the behaviour elsewhere → REMOVE. (Verify the
    paralog/organism before claiming this — see CLAUDE.md; do not REMOVE an
    experimental annotation just because the cached abstract foregrounds another
    gene.)
  2. Is the gene proximal to the behaviour? Neurotransmitter receptor,
    neuropeptide/hormone, ion channel, or circadian-clock component acting
    directly in the relevant circuit → behaviour may be near-core → ACCEPT
    or capture a more specific behaviour term.
  3. Is it a real but distal phenotype? (the common case — structural,
    metabolic, developmental, or ciliary gene whose knockout perturbs behaviour
    indirectly) → KEEP_AS_NON_CORE, with reason naming the proximal
    molecular defect the behaviour is downstream of.
  4. Is the term uselessly broad (adult behavior, behavior) or the
    phenotype barely connected?     → MARK_AS_OVER_ANNOTATED.

The default for a phenotype-driven behaviour annotation on a
molecular/structural gene is KEEP_AS_NON_CORE, not REMOVE: the phenotype
is genuine data about the gene, just not its core function.

Reproducing the analysis

uv run python projects/BEHAVIOR/mine_behavior.py \
    --genes-dir genes --out-dir projects/BEHAVIOR/reports

Outputs (regenerated, not hand-edited):

Spot-check of the ACCEPTed annotations

Applying the rubric to every behaviour annotation that a reviewer had ACCEPTed
as a core function sorts them cleanly into genuinely-proximal cases and missed
downgrades.

Genuinely proximal — ACCEPT upheld:

Missed downgrades — corrected to KEEP_AS_NON_CORE:

This moved 9 annotations from core to non-core, raising the downgrade rate among
adjudicated behaviour annotations in the June snapshot from ~81% to 87% (127 of 146; then only 19
ACCEPTed as core). The September 2026 re-run, over a larger corpus, gives 86%
(169 of 197; 28 accepted). Borderline cases left as-is (documented, not changed): daf-2
feeding/eating (the pleiotropic insulin receptor — feeding is one of many
outputs) and trpm7 swimming (a channel-kinase whose swimming phenotype is
plausibly a distal developmental consequence) — defensible either way and not
clear-cut enough to overturn.

ASSAY_TO_FUNCTION frames over-annotation risk on two
axes — proximity (does the readout measure the gene product's own molecular
activity, or a downstream cellular consequence?) and convergence (is the
readout a specific signature of process P, or a hub that many inputs feed into?).
A whole-animal behaviour is the maximal phenotypic + high-convergence readout:
it integrates the entire nervous system plus development, metabolism and basic
cell biology, so almost any perturbation can move it. That is exactly why 86% (169 of 197) of
adjudicated behaviour annotations are downgraded.

Behaviour has now been added as a first-class readout in that project's catalogue
(BEHAVIORAL_ASSAY in
readout_catalog.yaml), with the test
names — Morris Water Maze, open field, rotarod, fear conditioning, … — as match
patterns. The Casp3 swimming behavior case above is the emblematic failure
mode: the assay modality is mistaken for the gene's function. The Morris Water
Maze is a swimming-based test of spatial memory; a gene merely measured in it
(caspase-3, as an apoptosis marker) gets mis-annotated to swimming behavior.

Standardized behavioural-assay resources

There is no single canonical ontology that maps a behavioural assay to the GO
process it licenses (which is the gap readout_catalog.yaml fills by hand).
The landscape is split across three complementary layers:

For this project's purposes the practical takeaway is that an assay (IMPReSS/OBI)
reports a phenotype (MP/NBO), which is at best weak, non-core evidence for a GO
process (NBO/GO) — and never for a molecular function. Tightening the
behaviour branch's GO↔NBO alignment, and recording which assay drove each
behaviour annotation, would let the over-annotation check run automatically.

The IMPReSS standardized assay battery is ingested under
BEHAVIOR/impress/: a reproducible pull across 5
IMPC pipelines (287 procedures → 15 canonical behavioural/neurological assay
types, including Rotarod, Hole-board, Hot Plate, Tail Suspension, Von Frey and
Sleep-Wake, which the core pipeline omits), plus a hand-curated
behavioural_assay_go_map.yaml
mapping each assay to the GO behaviour term it can support as KEEP_AS_NON_CORE
(QuickGO-verified ids). That map closes the missing assay→GO link and fences off
the traps — Grip Strength (neuromuscular, no behaviour term), Tail Suspension (no
GO term for depression-like immobility) and Auditory Brain Stem Response
(electrophysiology, a hearing term at most, not auditory behavior).

The map is wired into the over-annotation mining two ways: a BEHAVIORAL_ASSAY
readout class in ASSAY_TO_FUNCTION/readout_catalog.yaml
(generic readout↔action cross-tab via mine_readouts.py), and a dedicated
check_behaviour_assays.py that
verifies the specific GO term against the specific assay named in an
annotation's evidence. The checker independently re-derived the Casp3 swimming behavior over-annotation (Morris Water Maze is a spatial-memory test; swimming
is only the modality) — confirming that fix from the assay side.

Status & next steps

Slides