GO-CAM curation & review

Use this skill to read and assess GO-CAM models. This project is read-only: we never edit or upload models. We cache them (gocams/<id>/<id>-src.yaml), index their activities (gocams/index.tsv), and record assessments in gocams/<id>/<id>-review.yaml (schema class GoCamReview).

What an activity (annoton) is

Each GO-CAM activity couples:

• enabled_by — the gene product carrying the activity;
• molecular_function — the GO MF term;
• part_of — optional biological process it contributes to;
• occurs_in — optional cellular component it happens in;
• causal_associations — directed edges to downstream activities.

In the cached YAML these are fields of each entry under activities:; they are flattened one-row-per-activity into gocams/index.tsv (join key: gene_product).

How to evaluate an activity

For each activity, work through these and record the outcome in the GoCamActivityReview:

1. Molecular function — most specific correct term; not a bare binding term; the real role (catalytic / receptor / adaptor / sequestering). → references/molecular-function.md
2. has input — names the substrate / target-gene / effector / cargo, not a ligand or raw DNA. → references/has-input.md
3. Causal edges — direct vs indirect, subject→object directionality, mechanism vs phenotype. → references/causal-relations.md
4. Complexes — right active subunit vs complex term. → references/complexes.md
5. Context & evidence — occurs_in, part_of, ECO + reference present.
6. Verdict, QC flags, and consistency with the gene review — the controlled vocabularies and the final checklist. → references/qc-and-consistency.md

Anchor every non-trivial verdict in verbatim supporting_text from a cited reference, the same discipline as gene-review supporting_text. Biological plausibility alone is not evidence.

Progressive disclosure

The references/ files hold the detailed rules; read the relevant one only when you hit that case. They map directly onto the schema enums (GoCamClaimVerdictEnum, GoCamQcFlagEnum, GoCamConsistencyEnum) so a review is machine-checkable:

• references/molecular-function.md — MF specificity, binding≠function, adaptor/sequestering/catalytic.
• references/has-input.md — has input semantics per activity type.
• references/causal-relations.md — direct/indirect causal, directionality, mechanism vs phenotype.
• references/complexes.md — complex subunit representation.
• references/qc-and-consistency.md — verdict scale, QC-flag glossary, consistency categories, checklist.

Workflow

just seed-gocam-review <model_id>      # one PENDING entry per cached annoton
# ... assess each activity using this skill ...
just validate-gocam-review gocams/<model_id>/<model_id>-review.yaml

The point of caching GO-CAMs here is the cross-check: join gocams/index.tsv to genes/**/<gene>-ai-review.yaml on gene product and record a GoCamConsistencyEnum verdict per activity. CONFLICT (the review removed / negated / over-annotated this function) and NOT_IN_REVIEW (a candidate gap) are the signals worth surfacing.

Attribution

Adapted from the GO Consortium GO-CAM annotation guidelines and the gocam-best-practice / validate-claims skills in geneontology/gocam-agent, tuned for this project's read-only ingestion and consistency-checking.