ai-gene-review ai-gene-review

active curating

ai-gene-review is an agent-curated knowledge base of Gene Ontology annotation reviews. One record per gene product: each record takes every existing GO annotation for the gene from GOA and gives it a review action (ACCEPT, MODIFY, REMOVE, MARK_AS_OVER_ANNOTATED, KEEP_AS_NON_CORE and others) with a reason and supporting evidence. The record also holds a standalone summary of the gene, its core functions as GO-CAM-like activities, proposed new GO terms, and suggested questions and experiments for experts. Reviews cover human, mouse, worm, yeast, fly, plant and many microbial gene products. Agents fetch the UniProt record, the GOA annotations and the cited publications, add deep research and bioinformatics, and write the review, which arrives by pull request. ai-gene-review is not strictly a Mech: each record reviews annotations that another resource holds, rather than being the canonical record for its gene. It keeps the Mech practices of a LinkML schema, ontology grounding, verbatim evidence and an append-only curation history.

One record is

a gene product and the review of its existing GO annotations

6760 records as of 2026-10-12

Counted from genes/*-ai-review.yaml at 00d0f98. Refreshed weekly.

Curation model

Agent-curated	Yes. AI agents generate or maintain most of the content.
Human review	`pull_request`. Agent changes arrive as pull requests that a person reviews and merges.
Agents and tools	Claude Code, dragon-ai-agent, Codex, goose
Evidence policy	Each review action carries a reason. Conclusive decisions on experimental annotations are expected to cite a PMID or DOI with supporting text quoted from the paper, and validation checks reference identifiers, titles and quotes against cached publication text. Ontology term identifiers are checked against their labels to catch fabricated terms.
Validation in CI	schema ontology_terms references
Append-only history	Yes
Record identifiers	Records are keyed by UniProtKB accession and named -ai-review.yaml under genes///. Organisms are a common name (human, mouse, worm, yeast) or a UniProt species code (ARATH, PSEPK). Gene symbols follow the nomenclature for the species.

Features

evidence_snippets Evidence items quote the cited source verbatim.
ontology_grounding Entities bind to public ontology terms with labels.
append_only_history Every change is an event in an append-only history directory.
knowledge_gaps Records carry structured open questions, controversies and curation to-dos.
static_browser A static site lets a reader browse every record.

Grounding and sources

Ontologies and vocabularies: GO ECO NCBITaxon

Source	Relation	What is taken
GOA	prov:hadPrimarySource	The existing GO annotations under review, downloaded through the QuickGO API.
UniProtKB	prov:hadPrimarySource	The UniProt record for each gene product, which supplies the record identifier.
PubMed	prov:used	Cited publications, cached as text so that supporting quotes can be checked.
Reactome	prov:used	Cached Reactome entries used as pathway context.
GO-CAM	prov:used	Cached production GO-CAM models, indexed by gene product.
PANTHER	prov:used	Protein family membership for family-level review and propagation.

Products

Product	Kind	Format	Description
Gene reviews `ai-gene-review.records`	RecordCorpus	yaml	One -ai-review.yaml file per gene product under genes///, next to the UniProt record, GOA annotations, notes and deep research it was written from.
Gene review LinkML schema `ai-gene-review.schema`	DataModelProduct	linkml linkml
Gene review browser `ai-gene-review.app`	GraphicalInterface	html	A searchable browser of the reviews. Review pages take thumbs-up and thumbs-down votes on individual decisions.
Exported annotations `ai-gene-review.annotations`	OtherProduct	csv	One row per reviewed annotation, with the review action, reason and supporting text.
Statistics report `ai-gene-review.stats`	DocumentationProduct	html

Relations to other Mechs

adopts_practice_of dismech: The history/ directory of append-only curation session records, with its schema, layout, scaffolder and validation, is ported from DisMech.

Contacts

AI4Curation (Organization) ai4curation
Christopher J. Mungall (Individual) ORCID 0000-0002-6601-2165 cjmungall@lbl.gov cmungall

Registry entry

Created 2026-10-06. Last modified 2026-10-06. Edit this entry on GitHub.

ai-gene-review is in the registry as a Mech-like resource. It shares the Mech curation practices, but a record is a review of annotations held elsewhere, in GOA, rather than the canonical record for its gene.

At commit 484ec6e on 2026-10-06, genes/ held 6,172 reviews across 210 organism folders. Human has the most at 2,656, then Pseudomonas putida KT2440 (PSEPK) at 955 and Arabidopsis thaliana at 257. The record count takes only *-ai-review.yaml files. The same folders hold 644 other YAML files, reviews of computational predictions and generated descriptions, which are not counted. history/ held 5,831 curation session records.

The repository also holds reviews of GO-CAM models, functional modules, Pfam and InterPro entries, and GO mappings for Rhea, TCDB, CAZy and other resources. These are outside the record count.