Co-Scientist, a hypothesis system, in Nature
On 19 May 2026 Nature published 'Accelerating scientific discovery with Co-Scientist' (Nature 655, 487–496), the peer-reviewed description of a multi-agent system built on Gemini that generates, critiques and ranks research hypotheses. The authors report laboratory tests, in cell-line experiments, of drug-repurposing candidates for acute myeloid leukaemia.
Why it matters
The claim that a system proposes hypotheses which experiment then supports now stands in a peer-reviewed journal, not only on a company blog. An editorial assessment: the experiments were run by the authors and their collaborators and the code is closed, so nobody outside has rerun the system.
What the article contains. Three biomedical applications: drug repurposing for acute myeloid leukaemia, novel epigenetic targets for liver fibrosis in human organoids, and a mechanism of bacterial gene transfer relevant to antimicrobial resistance. For the laboratory the experts chose three leukaemia candidates, nanvuranlat, KIRA6 and leflunomide; for KIRA6 a figure caption reports an 18-fold separation between the KG-1a cell line and the TK6 control. In the antimicrobial-resistance case the system was asked about a topic the group had already discovered but not yet published, and it proposed the top-ranked hypothesis independently. What the record does not claim. The article does not say a new drug was found: the testing was in vitro. The automated evaluation used 15 curated goals and the Elo analysis 203 goals entered up to 3 February 2025, so the system assessed is older than the publication (received 20 March 2025, accepted 11 May 2026). The article says the full source code is not public. The authors are at Google Cloud AI Research, Google DeepMind and Google Research together with Stanford University School of Medicine and others; no independent reproduction was read for this record.