Curate Geneset

Curating a GO interpretation of a gene set

You are adding (or reviewing) one curation/genesets/<SET>.yaml — a curator's GO-term interpretation of a non-GO gene set — to a hand-curated gold standard that doubles as a precision/recall benchmark for genesets-rs. One YAML per set.

Read first: curation/README.md (the categories, recovery_status, insight, sources, series, evidence — the authority) and the schema curation/schema/genesets_interpretation.yaml (slots + enums). Match the house style of an existing file, e.g. curation/genesets/TRAVAGLINI_LUNG_CILIATED_CELL.yaml.

The one rule that governs everything: ground in the actual membership

Before asserting any biology, read the set's real gene list. Do not curate from the cell type's textbook markers — curate from what is actually in the set. This is what makes recovery_status honest and is the most common way agents go wrong (e.g. a fetal "photoreceptor" set that lacks RHO/PDE6; a "chromaffin" set that lacks TH/DBH; a "pro-B" set dominated by cell-cycle genes).

How to get the membership:

• MSigDB sets (MSIGDB:<NAME>, collections C8/C2/H/…): _id on mygeneset.info equals the MSigDB set name. Fetch with scripts/fetch_mygeneset_query.py (or query _id:"<NAME>" directly). The full list is also what gets folded into the eval.
• Literature sets (LIT: ids): the defining paper is the identity, membership, and evidence. For short, explicit lists (GWAS loci, CRISPR hits, causal-gene panels) capture the genes verbatim from the primary source, HGNC-normalize, and append to curation/genesets/lit_members.gmt (tab-sep: <gene_set_name>\tcurated_literature_membership\tGENE1\tGENE2…). The GMT set_id MUST equal the YAML gene_set_name.

Per-term: the three orthogonal axes

For each association set category, confidence, specificity, seed_source, recovery_status, and (on core/supporting terms) insight, plus a curator_note and optional evidence. Aim for ~5–7 associations: the cell/state's defining core terms (a process + a component), 1–2 supporting, at most one nonspecific only if the membership clearly carries it (don't invent a housekeeping hit).

1. category — the biology (authoritative; NOT driven by annotation state). core_process / core_component / supporting_process / marker_driven_plausible / nonspecific / false_association. A term that is biologically core stays core even if no annotation supports it.

2. recovery_status — grounded in the actual membership (orthogonal to category). Name the specific present/absent genes in the curator_note:

   • annotation_supported — carrier genes are in the set; enrichment recovers it.
   • annotation_gap — relevant genes ARE in the set but GO annotation is too shallow. The gap is GO's (a GO-annotation curation target). seed_source: curator_added.
   • membership_gap — the carrier genes are NOT in the set (legacy/incomplete set, or a fetal signature lacking its mature effectors). The term still belongs in a complete description. seed_source: curator_added.

3. insight — interpretive value (only on core/supporting). The tightened rule:

   • confirmatory — entailed by the set's construction/identity (the default for a marker signature; ALL generic hallmarks — proliferation, apoptosis, generic PI3K/MAPK, known disease OXPHOS, expected immune terms).
   • mechanistic — a SPECIFIC, non-obvious process not entailed by construction; a genuine enrichment insight. Use sparingly (~10% of terms corpus-wide). When in doubt, it's confirmatory.

Anti-hallucination discipline (non-negotiable)

• OLS-verify every ontology id+label (GO/CL/UBERON/MONDO/…) before writing it. Use the OLS MCP (mcp__claude_ai_OLS__search / fetch). Confirm the id resolves to the exact label and is not obsolete. Never invent ids; if unsure, search by label and use the returned id.
• Reject obsolete terms. The validator's 4th gate sweeps OAK obsoletes() and fails the build. The sqlite:obo builds retain labels for obsolete classes, so an obsolete id paired with its old label passes id+label validation but fails here (e.g. GO:0062023 collagen-containing ECM → GO:0031012; GO:0050663 cytokine secretion → GO:0032635).
• Evidence snippets are verbatim or absent. Include an evidence item with a snippet ONLY if you retrieved the exact text from the cited paper (PubMed MCP) and copied it character-for-character — the reference-validator substring-checks it and fails on any mismatch. Otherwise use curator_note only. Never fabricate. Any item with a snippet must also carry a reference.
• Cite the right paper. Verify PMIDs. (The Descartes fetal atlas is PMID:33184181, not 32848094 — a recurring trap.)

Pairs / series

Link contrasting poles of one axis with series (a shared SERIES:<NAME> id) and series_role (this set's pole, free text). Add the field to both/all poles (edit the partner files too). The eval checks that opposite poles resolve to contrasting GO interpretations — e.g. SERIES:PANCREATIC_ISLET alpha/beta/delta, SERIES:MYELINATING_GLIA central/peripheral.

Validate, then regenerate the manifest

# 4 gates: linkml structural, term id+label, reference snippet, obsolescence
uv run --project python/genesets-workflows --extra curation \
  genesets-workflows curate validate curation/genesets/<SET>.yaml   # exit 0 = pass

Iterate until exit 0. Validate the whole corpus before a PR with just curate-validate.

Regenerate curation/genesets/manifest.tsv after a batch (columns: gene_set_id, collection, context_term, context_type, context_label, series, series_role) by looping every YAML through model.load_interpretation. Run just curate-test (29 unit + 5 doctest).

Fold the set into the eval (so it becomes scoreable)

A set is only evaluable if its membership is in the eval's queries.gmt. MSigDB sets come from the mygeneset base; LIT: sets come from lit_members.gmt. After adding sets, rebuild queries.gmt and re-run the evaluate-enrichment skill.

The guardrail (read evals/iba_vs_benchmark/README.md)

The eval measures the gold; it must never refit it. category is the scored truth. recovery_status/insight are curator judgments adjudicated against GOA facts during curation — never auto-updated to match whatever a tool happened to recover (that would make recall-vs-gold circular). When the eval surfaces a disagreement (a gap_recovered term), it is a review item, decided on the merits, not an automatic relabel.