Agentomics: an agentic system that autonomously develops novel state-of-the-art solutions for biomedical machine...

Area of Science:

• Biomedical data science
• Machine learning engineering
• Computational biology

Background:

• Automated machine learning (ML) is essential for advancing biomedical research and drug discovery.
• Current automated ML tools lack flexibility, and large language model (LLM) agents struggle with reproducible code generation.
• Existing LLM agent solutions underperform compared to human-engineered ML models.

Purpose of the Study:

• To introduce Agentomics, an autonomous LLM-powered agentic system for end-to-end ML experimentation in biomedicine.
• To enable the creation of flexible, reproducible, and high-performing ML models from biomedical datasets.
• To leverage biomedical foundation models within an automated experimentation framework.

Main Methods:

• Agentomics is an autonomous LLM-powered agentic system for end-to-end ML experimentation.
• It implements strict validation checkpoints for ML development, ensuring gradual, validated code progression.
• The system supports various LLMs and biomedical foundation models for diverse applications.

Main Results:

• Agentomics was evaluated on 20 datasets across Protein Engineering, Drug Discovery, and Regulatory Genomics.
• It outperformed other agentic systems in all tested domains.
• Agentomics generated novel state-of-the-art models for 11 out of 20 benchmark datasets, surpassing human expert solutions.

Conclusions:

• Agentomics provides a robust, autonomous solution for biomedical ML experimentation.
• The system demonstrates superior performance compared to existing automated methods and human experts.
• Agentomics facilitates the development of novel, state-of-the-art ML models for critical biomedical applications.