We're pleased to share the preprint of MetaboT: An LLM-based Multi-Agent Framework for Interactive Analysis of Mass Spectrometry Metabolomics Knowledge Graphs (Research Square).
MetaboT is a multi-agent system of large language models that lets researchers explore metabolomics knowledge graphs by simply asking questions in natural language. Behind the scenes, specialized agents translate a question into structured SPARQL queries over linked mass-spectrometry data, validate them against the graph schema, and return interpretable answers — making knowledge-graph-backed metabolomics accessible without query expertise.
On a 50-question benchmark, MetaboT reaches 83.67% accuracy, compared with 8.16% for a single LLM. The public demonstrator is connected to the Experimental Natural Products Knowledge Graph (ENPKG), built from a chemodiverse collection of 1,600 plant extracts.
MetaboT is a building block of the MetaboLinkAI consortium and the direct predecessor to our Mimosa framework for evolving multi-agent systems.
- Preprint: Research Square · 10.21203/rs.3.rs-6591884/v1
- Try it: metabot.holobiomicslab.eu
- Archive: Zenodo · 10.5281/zenodo.19715403
Bekbergenova M, Pradi L, Navet B, Tysinger E, Michel F, Feraud M, Taghzouti Y, Legrand M, Jiang T, Chen YZ, Hassoun S, Kirchhoffer O, Wolfender JL, Mehl F, Pagni M, Bittremieux W, Gandon F, Nothias LF.