Medea: An omics AI agent for therapeutic discovery.

Area of Science:

• Artificial Intelligence in Bioinformatics
• Computational Biology
• Genomics and Omics Data Analysis

Background:

• Artificial intelligence (AI) agents have the potential to accelerate biomedical discovery.
• Realizing this potential requires AI agents capable of transparent, long-horizon analyses using various tools.
• Ensuring transparency necessitates explicit intermediate decisions and validation against data and tool constraints.

Purpose of the Study:

• To introduce Medea, an AI agent designed for transparent, multi-step omics analyses.
• To enable AI agents to execute complex research objectives by integrating diverse tools and validation mechanisms.
• To enhance the performance and reliability of AI-driven biomedical research.

Main Methods:

• Medea integrates four modules: research planning with verification, code execution with checks, literature reasoning with evidence assessment, and a consensus stage.
• The agent utilizes 20 tools for transcriptomic data, cancer vulnerability maps, pathway knowledge bases, and machine learning models.
• Evaluations involved 5,679 analyses across target identification, synthetic lethality, and immunotherapy response prediction.

Main Results:

• Medea improved target identification by up to 46%, synthetic lethality reasoning by 22%, and immunotherapy response prediction by 24%.
• The agent demonstrated low failure rates and calibrated abstention across diverse evaluations.
• Performance improvements were observed across different large language models, tool sets, and agentic modules.

Conclusions:

• Verification-aware AI agents, like Medea, enhance biomedical discovery through transparent analyses, not just efficient workflows.
• Medea successfully reconciles evidence across datasets, tools, and literature for robust biomedical insights.
• The study highlights the importance of transparency and validation in advancing AI applications in omics research.