Evaluating agentic AI for biological discovery in autonomous and copilot settings

Area of Science:

• Computational Biology
• Artificial Intelligence
• Genomics

Background:

• Large language models (LLMs) enhance AI agents for structured bioinformatic pipelines.
• Biological discovery, especially in multi-omic studies, demands open-ended hypothesis generation and multimodal reasoning beyond deterministic pipelines.
• The utility of agentic AI in complex biological discovery remains largely unexplored.

Purpose of the Study:

• To systematically evaluate the capabilities and limitations of agentic AI in biological discovery.
• To assess AI agents' performance on multi-omic single-cell datasets across various cancer types.
• To establish a framework for evaluating AI systems supporting biological discovery.

Main Methods:

• Developed the Multistep Multimodal Multiomic Agentic (M3A) Framework for LLM-driven reasoning over multimodal data.
• Assessed AI agents on tasks including autonomous cell-type annotation and hypothesis generation.
• Conducted copilot experiments to evaluate human involvement and domain expertise.

Main Results:

• AI agents demonstrate effectiveness in broad, systemic exploration of complex multi-omic data.
• Domain experts are crucial for methodological guidance and biological synthesis.
• Current AI agents show potential but have limitations in complex biological reasoning.

Conclusions:

• Agentic AI shows promise for exploring complex biological datasets.
• Human expertise remains indispensable for interpreting results and advancing biological discovery.
• The M3A framework provides a standardized method for assessing AI in computational biology.