Open-source AI co-scientist completes expert-reviewed biomedical analyses in minutes, its developers report
Nature Medicine describes OpenScientist, an auditable agentic assistant evaluated by domain experts on five biomedical case studies.
A paper published in Nature Medicine on 16 November 2026 introduces OpenScientist, an open-source agentic system designed to investigate questions set by scientists and to produce analyses that can be checked against the underlying data. The work comes from Washington University in St. Louis, Lawrence Berkeley National Laboratory, and collaborators, with co-first authors Kaleigh F. Roberts and Zachary B. Abrams. The article is released under a Creative Commons Attribution 4.0 license.
Domain experts evaluated the software on five biomedical case studies. The paper's abstract reports that the system quantified the performance of Alzheimer's blood biomarkers, built 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 that were validated in an external cohort, with a weaker signal in randomized controls, and agreed with an independent human analysis in a blinded, prespecified proteomic comparison.
The authors write that OpenScientist completed those analyses "in minutes that otherwise would take months of human time and expertise." That timing is their comparison, not the result of an independent time-and-motion study. The evaluation covers five expert-scored case studies. The authors present the software as an assistant that works with human scientists, not as a system that makes discoveries on its own.
For research that involves sensitive clinical data, they say OpenScientist can run in a fully air-gapped setting, using open-source models for inference. That arrangement is meant to keep patient data off external networks. The software is open source, and its code is public on GitHub.
Why it matters
Closed systems are hard to audit from the outside. A peer-reviewed, open system that can run without an internet connection offers a more transparent way to test whether agentic tools actually help with real biomedical datasets — a practical requirement when the data cannot leave a secure lab.
Looking ahead (speculative)
If independent groups can reproduce the five case studies and try the software on new datasets, tools of this kind could shorten the early grind of analysis. They will still need expert review, and it is too soon to know how well the approach will hold up outside the studies reported here.