IMPReSS behavioural-assay ingest + assay→GO check

Standardized phenotyping assays from IMPReSS (the International Mouse
Phenotyping Resource of Standardised Screens, IMPC), ingested to operationalize
the BEHAVIOR ↔ ASSAY_TO_FUNCTION
cross-link: which standardized assay drives a behaviour annotation, and what GO
process (if any) it can legitimately support.

Files

• ingest_impress.py — reproducible fetch of one or more pipelines' full
  procedure sets from the IMPC API (https://api.mousephenotype.org/impress),
  classifying behavioural/neurological procedures into canonical assay types.
  Live fetch; nothing hard-coded.
• procedures.tsv (generated) — every procedure across the ingested
  pipelines (pipeline, key, name, level, version, #parameters, MP terms,
  behavioural flag, canonical assay type).
• behavioural_assays.yaml (generated) — behavioural/neurological assays
  grouped by canonical type, each listing every (pipeline, key) instance.
• REPORT.md (generated) — counts + the canonical behavioural-assay table.
• behavioural_assay_go_map.yaml (hand-curated) — the bridge: each
  behavioural assay type → the GO behaviour term(s) a hit can support, as
  KEEP_AS_NON_CORE (all ids QuickGO-verified), with the over-annotation
  caution. assay_type joins to behavioural_assays.yaml. Includes the
  Morris Water Maze (flagged not_in_impress: a classic literature assay).
• check_behaviour_assays.py + ASSAY_CHECK.md / assay_check_flags.csv
  (generated) — scans every annotation whose evidence text names a
  recognised assay and flags MISMATCH (assay doesn't license the GO term),
  OVER_CORE (accepted as core), or FENCED (assay licenses no behaviour term).
• validate_map.py — validates behavioural_assay_go_map.yaml (these files
  are bespoke project data, not bound to the LinkML gene-review schema, so this
  is their dedicated check). Offline: parse, structure, unique assay_type,
  well-formed GO:nnnnnnn/label entries, and an exact join to the ingested
  canonical assay types. --online: every GO id resolves in QuickGO, is
  non-obsolete, is a biological_process, and its label matches verbatim
  (checked per occurrence). Exits non-zero on failure.

bash uv run python projects/BEHAVIOR/impress/validate_map.py --online

Regenerate

A project-local justfile wraps the workflow (recipes cd to the repo root):

cd projects/BEHAVIOR/impress
just                 # list recipes
just ingest          # refresh procedures.tsv / behavioural_assays.yaml / REPORT.md
just validate-online # structure + exact join + QuickGO id/label check
just check           # ASSAY_CHECK.md / assay_check_flags.csv
just all             # ingest -> validate-online -> check

Or call the scripts directly from the repo root:

uv run python projects/BEHAVIOR/impress/ingest_impress.py --out-dir projects/BEHAVIOR/impress
uv run python projects/BEHAVIOR/impress/check_behaviour_assays.py --out-dir projects/BEHAVIOR/impress
uv run python projects/BEHAVIOR/impress/validate_map.py --online

The default ingest covers 15 canonical behavioural/neurological assay types
across 5 pipelines — including the assays absent from the IMPC core pipeline
(Rotarod, Hole-board, Hot Plate, Tail Suspension, Von Frey, Sleep-Wake). Pass
--pipelines to ingest others.

How the check wires into the readout mining

Two complementary layers:

1. Generic — BEHAVIORAL_ASSAY is a readout class in
   ASSAY_TO_FUNCTION/readout_catalog.yaml, so mine_readouts.py already
   cross-tabulates behavioural-assay mentions against reviewer action.
2. Specific — check_behaviour_assays.py uses
   behavioural_assay_go_map.yaml to verify the exact GO term against the
   specific assay. This independently re-derived the Casp3 swimming behavior → Morris Water Maze over-annotation (the assay is a spatial-memory
   test; swimming is only the modality).

The check is exact-id and advisory: it matches the annotated term id against
the assay's licensed set (a future refinement would use GO subsumption so child
terms pass automatically), and OVER_CORE should be read as "confirm this gene
is genuinely proximal" — bona-fide clock genes / receptors are the documented
rubric exceptions (e.g. CRY sleep/wake is correctly near-core).

Why a hand-curated GO map

No public ontology maps "assay → GO process": NBO / MP
(behavior/neurological phenotype, MP:0005386) describe the process/phenotype,
IMPReSS (this ingest) / OBI the assay, CogPO / Cognitive Atlas
human cognitive tasks. behavioural_assay_go_map.yaml closes that gap for the
IMPReSS battery, with the ASSAY_TO_FUNCTION caution baked in (phenotypic +
high-convergence → non-core only). Two assays are deliberately fenced (empty
supports_go): Grip Strength (neuromuscular) and Tail Suspension (no GO
term for depression-like immobility); Auditory Brain Stem Response is capped
at a hearing term and fenced against auditory behavior.

Caveats

• MP terms come from the procedure description; richer parameter-level MP
  mappings are available via /ontologyterm/belongingtoparameter/:parameterID.
• The checker only sees annotations that name an assay in their evidence text;
  most cite a PMID instead, so its flag count is a floor, not the full
  population.