Improving package annotation in metabolomics and proteomics via robust, ontology-driven LLM integration

Identifying the most appropriate bioinformatics tool for a task remains challenging across multiple domains. Annotating tools with EDAM ontology terms (e.g. topics, operations, input / output data and formats) can help, but manual annotation is labour-intensive, error-prone, and difficult to scale, particularly given the high rate of first-time package developers in academic environments. At BioHackathon Europe 2025, our team explored how Large Language Models (LLMs) can assist this process through the Model Context Protocol (MCP), an emerging open standard that specifies how LLMs call external functions, using metabolomics as a domain use case. We developed an MCP-based workflow that grounds tool descriptions in the EDAM ontology (Ison et al., 2013), improving reproducibility and semantic precision. Two core modules, entry-point specification and semantic text segmentation, were completed during the hackathon, while additional mapping, validation, and reporting functions were outlined for follow-up development. Benchmarking integrated with the BioChatter framework (Lobentanzer et al., 2025) demonstrated that MCP-assisted models outperform unconstrained baselines on initial tests using metabolomics packages from bio.tools (Ison et al., 2019). Ongoing work will expand benchmarking datasets, refine term-mapping logic, and extend the workflow to proteomics, supporting scalable, ontology-driven annotation across the ELIXIR ecosystem.