Empowering AI data scientists using a multi-agent LLM framework with self-evolving capabilities for autonomous,...

Area of Science:

• Biomedical Data Analysis
• Artificial Intelligence in Science
• Bioinformatics Tool Integration

Background:

• Large language models (LLMs) offer potential for automating complex tasks and scientific data exploration.
• Current LLM applications in biomedical data analysis are constrained by challenges in handling specialized tools and multistep reasoning.
• There is a need for advanced AI frameworks that can effectively integrate diverse bioinformatics tools for complex biomedical research.

Purpose of the Study:

• To introduce BioMedAgent, a self-evolving LLM multi-agent framework designed for biomedical data analysis.
• To enable users to initiate complex tasks using natural language, bypassing the need for computational expertise.
• To develop a system capable of learning to use and chain diverse bioinformatics tools into executable workflows.

Main Methods:

• Developed a self-evolving LLM multi-agent framework named BioMedAgent.
• Implemented interactive exploration and memory retrieval algorithms for tool utilization and workflow chaining.
• Trained and evaluated BioMedAgent on the BioMed-AQA benchmark (327 tasks) and the external BixBench dataset.

Main Results:

• BioMedAgent achieved a 77% success rate on the BioMed-AQA benchmark, outperforming other LLM agents.
• Demonstrated robust generalization capabilities on the external BixBench dataset.
• Successfully performed autonomous cross-omics analysis, machine-learning modeling, and pathology image segmentation.

Conclusions:

• BioMedAgent represents a significant advancement in applying LLMs to biomedical data analysis, overcoming limitations in tool handling and multistep reasoning.
• The framework empowers biomedical researchers by enabling natural language task initiation and automating complex computational workflows.
• BioMedAgent shows broad potential for advancing biomedical research and can be extended to other scientific domains requiring sophisticated tool integration.