Nature Medicine has published OpenScientist, an open-source agentic AI “co-scientist” for biomedical research, in an open-access paper dated 16 November 2026.
The work comes from a collaboration led by researchers at Washington University in St. Louis, the Consortium for Biomedical Research & AI in Neurodegeneration (C-BRAIN) and Lawrence Berkeley National Laboratory, with co-first authors Kaleigh F. Roberts and Zachary B. Abrams. Other co-authors include researchers at Harvard Medical School, the UK Dementia Research Institute at University College London, and further institutions in the United States and Europe. The paper is released under a Creative Commons Attribution 4.0 International licence.
The authors describe OpenScientist as “an open-source agentic AI co-scientist designed to address this gap by investigating scientist-defined queries and generating verifiable, data-grounded analyses.” They say it was evaluated “by domain experts across five biomedical case studies, completing analyses in minutes that otherwise would take months of human time and expertise.”
What the OpenScientist paper found in five case studies
According to the paper, OpenScientist “accurately quantified Alzheimer’s blood biomarker performance, generated a plasma proteomic survival model that predicted survival in a hidden-evaluation cohort, proposed a reproducible mechanism linking tau pathology to altered lysosomal acidification, generated multiple myeloma hypotheses validated in an external cohort with attenuated signal in randomized controls, and demonstrated concordance with an independent human analysis in a blinded, prespecified proteomic comparison.”
Those five studies used existing datasets. The authors stop short of presenting OpenScientist as a tool for clinical use on its own. They write: “These results support the use of OpenScientist to assist biomedical data analysis in concert with human experts.”
To support work on sensitive clinical datasets, they also say “OpenScientist can operate in a fully air-gapped environment using open-source models for inference.” That is a stated capability of the software, not a report of a hospital deployment.
The paper argues that OpenScientist can support real-world clinical research by researching the literature, generating hypotheses, executing analyses and generating insights from complex datasets. That closing claim is the authors’ description of what the system is designed to do, not a finding from a prospective clinical trial.
Code, data and the earlier preprint
The authors have released the code on GitHub and a supporting data deposit on Zenodo. An earlier version of the work appeared as a medRxiv preprint in March 2026. The project’s GitHub repository also lists documentation at docs.openscientist.io.
TSN has previously covered other attempts to put AI agents into laboratory science, including Xaira’s claim that its X-Scientist system matched human scientists at picking gene targets and a later report that Xaira’s X-Bot laboratory arm kept a lab instrument 97% busy for a week. Those efforts sit in a wider turn towards teams of AI agents that split up scientific or operational work, of which OpenScientist is one open-source, peer-reviewed example aimed at biomedical data analysis rather than a physical lab robot.
A co-scientist that works in minutes is still an assistant
The headline finding that OpenScientist can complete in minutes analyses that would otherwise take months comes from expert evaluation of five case studies, not from a head-to-head clinical trial or a time-and-motion study of a working hospital laboratory. The paper is clear that the system is meant to assist “in concert with human experts.” That framing matters, because the five results are not all the same kind of claim.
Quantifying the performance of Alzheimer’s blood biomarkers, and showing that a plasma proteomic survival model predicted survival in a hidden-evaluation cohort, are retrospective analyses of data that already existed. A “proposed” mechanism linking tau pathology to altered lysosomal acidification remains a proposed mechanism until it is independently tested. The authors do say that multiple myeloma hypotheses were “validated in an external cohort,” with an attenuated signal in randomised controls; that is their wording for a statistical check on existing data, not a licence to treat those hypotheses as clinical findings. Concordance with an independent human analysis in a blinded, prespecified proteomic comparison is encouraging, and it is still a comparison on a prepared dataset.
The air-gapped, open-source-model mode is a genuine practical point for laboratories that cannot send clinical records to a commercial cloud. It does not, by itself, answer questions about model quality, audit trails or who is responsible when an agent’s report is wrong. OpenScientist is now a peer-reviewed, openly licensed tool that domain experts have put through five biomedical case studies. The next test is whether other groups can reproduce those analyses, and whether the same “minutes versus months” claim holds when the questions and the datasets are new.
Sources
- Nature Medicine, “An open agentic AI co-scientist for clinical and biomedical discovery”, 16 November 2026
- OpenScientist source code, GitHub
- OpenScientist data deposit, Zenodo
- OpenScientist medRxiv preprint, March 2026
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterRelated stories
- Xaira Says Its X-Scientist AI Matched Human Scientists at Picking Gene Targets is the closest recent TSN file on an AI system being compared with human scientists on a biomedical task, and the right next stop for that contrast.
- Xaira Says Its X-Bot Arm Kept a Lab Instrument 97% Busy for a Week
- What Are AI Agent Swarms? How Teams of AI Agents Split Up the Work
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
