Louis-Felix Nothias - HolobiomicsLab

Today at #MetSoc2026 I gave a keynote presentation on self-evolving, multi-agent AI for metabolomics — from @holobiomicslab.bsky.social. The headline outcome: we're releasing the first ASB Skill Collection — open, evidence-grounded metabolomics building blocks for any scientific AI agents. 🧵👇

The source corpus builds on the computational-metabolomics review series started by @biswamisra.bsky.social (2014) @jjjvanderhooft.bsky.social , now maintained by Enveda — thanks to Tobias Kind for great team work in Dagstuhl. 11/

doi.orgASB Metabolomics Skill Collection v2 (metabolomics-v0.1.0)Curated, evidence-grounded skill and software-tool collection for computational metabolomics, generated by the AgenticScienceBuilder (ASB) pipeline: 5,865 skills across 909 tools distilled from 568 pe...

Thanks to the ASB community & Dagstuhl participants. API costs sponsored by @cnrs.fr & @univcotedazur.bsky.social . Built @holobiomicslab.bsky.social (eDIAM team,with Élise Dumont) with @metabolinkai.bsky.social . Zenodo: doi.org/10.5281/zeno... 12/12

github.comGitHub - HolobiomicsLab/asb-skill-collections: Curated, evidence-grounded skill and software-tool collections for scientific AI agents, generated by the AgenticScienceBuilderCurated, evidence-grounded skill and software-tool collections for scientific AI agents, generated by the AgenticScienceBuilder - HolobiomicsLab/asb-skill-collections

Try it in Claude Code in two lines: /plugin marketplace add HolobiomicsLab/asb-skill-collections /plugin install metabolomics@asb-skill-collections Repo + other runtimes: github.com/HolobiomicsL... 9/

My take-home: the bottleneck left is not technical — it's collective. The community must validate, curate & govern these benchmarks and skills transparently, so science, not the vendor, stays in control. Try it, break it, join us and contribute to the upcoming community paper. 10/

So as a community effort started at Dagstuhl, we built the AgenticScienceBuilder (ASB): a pipeline that reads real method papers and distills them into benchmark tasks, skills & tools — each verified against its source. 7/

👉 The release: the first ASB Skill Collection for metabolomics. 🔹 5,865 skills · 909 tools · 568 papers 🔹 LC-MS/MS + GC-MS, MS-imaging, ion mobility, lipidomics, NMR 🔹 evidence-grounded, EDAM-annotated, open (CC-BY-4.0) 8/

Step 3: keep agents honest. Perspicacité grounds them in the literature — an open, local-first literature-RAG engine where the evidence is yours and traceable, not a black box. github.com/HolobiomicsL... 5/

But is an agent good at our science? That question was sharpened at the Dagstuhl Seminar on Computational Metabolomics (April 2026). The bottleneck: no metabolomics benchmark, no domain skills/grounding. 6/

doi.orgMetaboT: An LLM-based Multi-Agent Framework for Interactive Analysis of Mass Spectrometry Metabolomics KnowledgeMass spectrometry metabolomics generates complex data that overwhelms traditional analysis approaches. MetaboT, a multi-agent Large Language Model (LLM)-based framework, converts natural-language ques...

Step 1: make the data queryable. MetaboT is a multi-agent system that turns plain-language questions into validated SPARQL over a metabolomics knowledge graph, tested on 50 expert questions. DOI: doi.org/10.21203/rs.... 3/

doi.orgMimosa Framework: Toward Evolving Multi-Agent Systems for Scientific ResearchCurrent Autonomous Scientific Research (ASR) systems, despite leveraging large language models (LLMs) and agentic architectures, remain constrained by fixed workflows and toolsets that prevent adaptat...

Step 2: make the agents better. Mimosa is a self-evolving, auditable system that reconfigures itself from feedback — 43% on ScienceAgentBench (102 expert tasks), beating the prior best at lower cost. arXiv: doi.org/10.48550/arX... 4/

The community has the ability and tools for producing metabolomics data at scale — a single plant/microbe collection can hold 10,000+ extracts and millions of spectra. We're not yet as good at asking it everything we could. The answer lives across spectra, taxa, annotations & assays at once. 2/

linkedin.com#wageningen #dagstuhl #proudpi #compmetabolomics #metabolomics #bioinformatics | Justin J.J. van der HooftGetting closer to #Wageningen after an inspiring #Dagstuhl Computational Metabolomics seminar! 😎 It was a week full of agents that churn, marinate, vibe, and what is more, benchmarks, substructures, ...

I love the photos were the colleagues are checking out the preliminary results of the prototype for the benchmark pipeline on my computer (the one with a CNRS sticker!) during one legendary "beer session".

Justin J.J. van der Hooft@jjjvanderhooft.bsky.social

www.linkedin.com/posts/jjjvan...

Thanks and congratulations to @jjjvanderhooft.bsky.social, Soha Hassoun @pluskal-lab.org, Stacey Reinke for organizing this landmark #Dagstuhl Computational Metabolomics seminar 26181 ! www.dagstuhl.de/en/seminars/... What an amazing four days with many fruitful discussions and interactions ...

arxiv.orgMimosa Framework: Toward Evolving Multi-Agent Systems for Scientific ResearchCurrent Autonomous Scientific Research (ASR) systems, despite leveraging large language models (LLMs) and agentic architectures, remain constrained by fixed workflows and toolsets that prevent adaptat...

... like Mimosa-AI (arxiv.org/abs/2603.28986) and those from MetaboLinkAI More generally, I was pleased to notice the general interest in agentic-AI and admission of its transformative potential for metabolomics and science - which wasn't that widely recognized 2 years ago.

At the same time, this was very important to have discussions with colleagues on the short/long-term challenges/risks. There are many and we should anticipate.

This time, in addition to act as a notetaker for several sessions, I co-moderated, with Florian Huber and Tobias Kind, a session on "Benchmarking tasks/workflows for metabolomics agent systems" on 29th April 2026.

The first sessions on AI converged to the existing gap, and that follow-up session enabled some powerful brainstorming with participants. We are now crafting an aligned benchmark that will help in designing and evaluating emerging agentic AI systems for metabolomics, ...

... meeting old and new friends. I am so grateful to be part of seminar (again), which truly shaped our community and research direction over the year.

Louis-Felix Nothias - HolobiomicsLab reposted

We've been building multi-agent AI for science at @holobiomicslab.bsky.social. We just released Mimosa (preprint, 30 March 2026) — a framework where the multi-agent workflows themselves evolve instead of following a fixed recipe. arxiv.org/abs/2603.28986

github.comGitHub - HolobiomicsLab/Mimosa-AI: Open-source Evolving AI-Framework for Autonomous Scientific Research (ASR). Evolve multi-agent workflows to perform computational scientific tasks.Open-source Evolving AI-Framework for Autonomous Scientific Research (ASR). Evolve multi-agent workflows to perform computational scientific tasks. - HolobiomicsLab/Mimosa-AI

Both Mimosa and Toolomics (MCP server management) are open source under Apache 2.0. Model-agnostic. Designed for community extension. Code: github.com/HolobiomicsLab/Mimosa-AI Toolomics: github.com/HolobiomicsLab/Toolomics DOI: doi.org/10.48550/arXiv.2603.28986

Built at @holobiomicslab.bsky.social sky.social @cnrs.fr @univcotedazur.bsky.social/ @icn-cnrs.bsky.social / 3iA Côte d'Azur), with Inria Wimmics @fabien-gandon.bsky.social . Within @metabolinkai.bsky.social — funded by @agencerecherche.bsky.social & @snsf.ch . /cc @cnrscotedazur.bsky.social

Why does academia need to build its own AI tools? If we stop building because big tech is ahead, we lose the skills and infrastructure to govern the tools research will depend on. Openness isn't just principle — it's how science retains its sovereignty.

Next steps for Mimosa: – Population-based workflow exploration – Domain-specific skill acquisition from published methods – Persistent failure-aware memory – Evaluation on lightweight open-source models Goal: a community-extensible engine for autonomous scientific AI.

On ScienceAgentBench (102 real scientific tasks, 4 disciplines): – Single agent: 38.2% – Static multi-agent: 32.4% – Iterative learning: 43.1% Tested across bioinformatics, computational chemistry, GIS, and cognitive neuroscience.

Everything is saved: every workflow topology, every agent prompt, every intermediate file. Tools run in containerized MCP servers — environment frozen, dependencies locked. A Python script can replay any step. Fully logged. Auditable. Replayable.

Most AI systems for science follow fixed pipelines. But research is non-linear — you screen, dock, simulate, realize your assumptions were wrong, and restart. Fixed architectures can't adapt. Mimosa was built to change that.