AI Gene Review
A review of existing Gene Ontology annotations, one file per gene, with a public voting interface for feedback.
Repository: ai4curation/ai-gene-review
App: Browse gene reviews
Product: genes/<organism>/<GENE>/<GENE>-ai-review.yaml
Agent surface: local sessions, GitHub mention, pull request review
What to look at
| Path | What it is |
|---|---|
.claude/skills/ |
Fifteen skills, including annotation-reviewer, core-function-synthesizer, gocam-curation, and aigr-pr-review |
.claude/hooks/ |
Hooks that run around agent actions |
genes/ |
One directory per gene, holding the review and its cached references |
src/ai_gene_review/schema/ |
The LinkML schema |
What works
Every annotation gets a verdict and a reason
A review does not rewrite an annotation. It records an action against the
existing one, such as ACCEPT, MODIFY, or REMOVE, together with the reason:
existing_annotations:
- term:
id: GO:0005515
label: protein binding
action: MODIFY
reason: "While evidence is strong, 'protein binding' is uninformative..."
This is reviewable in a way that a rewritten file is not. A curator reads the verdict and the reason, and does not have to reconstruct what changed.
Identifiers carry their labels
Every term appears as an identifier and a label together. A wrong pair fails validation, because an agent would have to fabricate both consistently to get past the check. See Make identifiers hard to fake.
Non-curators can give feedback
The generated pages carry thumbs-up and thumbs-down controls, and there is a longer evaluation form for detailed review. A domain expert can disagree with an agent without learning Git.
This is the part most projects skip. Validation catches fabricated evidence. Only a human who knows the gene catches a claim that is well-cited and wrong.
Skills split reviewing from synthesizing
annotation-reviewer judges what exists. core-function-synthesizer writes what
the gene does. aigr-pr-review reviews the pull request. Three jobs, three
skills, three sets of instructions that do not interfere.
What to copy first
Copy the action-and-reason record shape. If your agents change existing curated statements, record the verdict and the reason next to the original instead of replacing it. Review gets much cheaper.
Gaps
The public voting data is feedback, not yet a measurement. Turning votes into a number you can track over releases is still open work, and it is the obvious next step for anyone copying this pattern.