¸£ÀûÔÚÏßÃâ·Ñ

DSP Seminar: Agentomics: Autonomous Machine Learning Experimentation with LLM Agents

Advances in biomedical research increasingly rely on Machine Learning (ML) to extract insights from growing biological datasets. However, developing and evaluating effective ML models requires significant engineering expertise, creating a gap between modern ML methods and their practical use by experimental biologists. Recent advances in Large Language Models (LLMs) have introduced the possibility of autonomous AI agents that can assist in scientific workflows, including the design and implementation of ML pipelines.

In this talk, we explore the emerging paradigm of agentic systems for scientific ML. We present Agentomics, an autonomous LLM-powered system designed to perform end-to-end ML experimentation for biomedical datasets. Given a dataset, Agentomics iteratively explores modelling strategies, generates training pipelines, evaluates models, and produces reusable artifacts for inference and retraining. The system has been evaluated across multiple biomedical domains, including Protein Engineering, Drug Discovery, and Regulatory Genomics. In several benchmark datasets, Agentomics discovers models that surpass state-of-the-art human-engineered solutions, demonstrating the potential of agentic systems to automate complex ML experimentation.

This is the seventh talk in the 2025/6 Data Science Platform Seminar Series.

Speaker’s Bio

Andrea Gariboldi is a Research Support Officer at the University of Malta and a member of the Bioinformatics for Genomics in Malta (BioGeMT) research team. His work focuses on the development of Large Language Model agents for automating supervised machine learning pipelines. His research also explores scalable database architectures for the efficient storage, management, and retrieval of genomic variant data, supporting the development of robust computational infrastructure for genomics research.