Using agentic AI to assign gene function across multiple genomes

The Gene Ontology (GO) provides a foundational framework for representing gene function across species, encompassing a structured vocabulary, gene function assertions, and causal activity models (GO-CAMs). Maintaining and extending this knowledge base requires expert biocurators to distill findings from the scientific literature—a process that cannot scale to match the pace of modern genomics. While machine learning approaches to GO function prediction have attracted significant interest (e.g., the CAFA challenge series), rigorous evaluation reveals persistent limitations, particularly for uncharacterized proteins.

Agentic AI represents a paradigm shift from earlier "LLM-as-oracle" approaches that relied on narrow context windows and were prone to hallucination. Modern agentic systems combine a simple but powerful architecture—loop, tools, and state—enabled by large context windows, reasoning-capable models, and universal coding agents (e.g., Claude Code, Codex). Through the Model Context Protocol (MCP) and agent skills, these systems achieve a decoupled architecture that can autonomously plan, research, and execute complex curation tasks while maintaining a collaboration dial between full autonomy and human oversight.

We describe our deployment of multiple agentic AI systems for GO curation: an Ontology Agent for vocabulary maintenance, a Gene Annotation Review Agent that can re-review hundreds of annotations and flag biologically incorrect automated predictions, a GO-CAM Pathway Agent for constructing causal models, and OpenScientist, a co-scientist agent for interpreting molecular data. To combat hallucinations—particularly of ontology term IDs, citations, and clinical codes—we employ LinkML-based agentic harnesses with expressive but tightly constrained schemas, anti-hallucination checksums, and dedicated term and reference validators.

We frame human-agent collaboration as a pyramid: deterministic processes (automated QA/QC, schema validation) at the base, defined high-level agent processes (research, report generation, ontology maintenance) in the middle, and human management and governance at the top—shifting the role of humans from being "in the loop" to "regulating the loop." Evaluation remains challenging given the jagged frontier of AI capabilities, but validation experiments with OpenScientist demonstrate that the platform can distinguish true biological signals from technical noise. By automating lower-level curation tasks, agentic AI frees biocurators to focus on high-level modeling, consensus building, and direct engagement with the research community.