Multi-agent AI systems for autonomous science

A defining focus of the lab is building open multi-agent AI systems for science. We design agents that can autonomously operate and optimise mass spectrometers, query metabolomics knowledge graphs in natural language, and orchestrate end-to-end data-driven discovery. Our Mimosa framework takes this furthest — automatically generating, executing, and iteratively refining task-specific multi-agent workflows, a step toward next-generation autonomous scientific research — supported by MetaboT (a multi-agent LLM for metabolomics) and the Toolomics MCP tooling platform.

Alongside the agentic stack we train foundational mass-spectrometry models, leverage large language models for hypothesis generation and literature mining, and build human–computer interaction layers that keep scientists in the loop.

Research aims

We are committed to creating intuitive, open AI-powered frameworks that expedite molecular discovery and foster data-driven insights. By harnessing state-of-the-art computational advances — including natural-language processing for hypothesis generation, dynamic visualization tools, and interactive analytical frameworks — our primary goal is to transform the interpretation of metabolomics data through an open-access metabolomics AI assistant, designed to extend seamlessly to integrative multi-omics applications.

Current projects

• MetaboLinkAI — "Open, Integrative, and Extendable AI and Knowledge Graphs Framework for Functional and Actionable Metabolomics" (2025–2029), co-funded by the Swiss National Science Foundation (SNF 10002786, €3.16M) and the French ANR (ANR-24-CE93-0012-01, €1.3M). Coordinated by ETH Zurich and CNRS, uniting eight institutions across France and Switzerland. The HolobiomicsLab is co-coordinator and leads development of the metabolomics AI assistant. metabolinkai.net
• KGbot — Knowledge-Graph Chatbot for Metabolomics. Uses language models to translate natural-language queries into structured requests over knowledge graphs. Funded by Académie 1, IdEx Université Côte d'Azur (€70k, 2024–2025) with support from 3iA Côte d'Azur.
• Deep Reinforcement Learning for MS data acquisition (Madina Bekbergenova, PhD). Employs offline RL to dynamically adjust instrument configurations in real time, maximising the quality and coverage of fragmentation spectra. Part of an international PhD with Prof. Wout Bittremieux, funded by EUR Spectrum and 3iA Côte d'Azur (2024–2027, €120k).

Selected publications

• Tysinger, E., Pagni, M., Kirchhoffer, O., et al. (2023). An artificial-intelligence agent for navigating knowledge-graph experimental metabolomics data. Swiss Metabolomics Society Annual Meeting. hal-04381448
• Schmid, R., Petras, D., Nothias, L.-F., et al. (2021). Ion-identity molecular networking for mass-spectrometry-based metabolomics in the GNPS environment. Nature Communications, 12, 3832. 10.1038/s41467-021-23953-9
• Nothias, L.-F., Petras, D., Schmid, R., et al. (2020). Feature-based molecular networking in the GNPS analysis environment. Nature Methods, 17, 905–908. 10.1038/s41592-020-0933-6