DeepMind unveils Co-Scientist, an AI research partner that already helped find a liver-disease drug candidate
Google DeepMind described Co-Scientist, a Gemini-based multi-agent system that generates and tests scientific hypotheses, after a Stanford team used it to surface a drug candidate that blocked 91 percent of a scarring response in lab tests.
Artificial IntelligenceGNGV Editorial Team5 min read
Google DeepMind described a system called Co-Scientist, a multi-agent artificial intelligence built on its Gemini models that is designed to work alongside researchers by generating, debating and refining scientific hypotheses. Rather than answering a single question, the system runs several specialized AI agents that propose ideas, critique one another and rank the most promising directions, with the goal of accelerating the slow early stages of discovery. The work was reported in the journal Nature on May 19, 2026.
The most concrete early result came from medicine. A team led by Gary Peltz at Stanford University School of Medicine used Co-Scientist to look for existing, already-approved drugs that might be repurposed to treat liver fibrosis, the scarring that drives chronic liver disease. The system highlighted overlooked candidates, and in laboratory tests one of them blocked 91 percent of a scarring-linked response, a striking result the team published separately in Advanced Science.
The tool has been applied beyond the liver. DeepMind and its collaborators reported uses in antimicrobial-resistance research, plant immunity and the biology of cellular aging, where Omar Abudayyeh's lab said the system cut some analysis from months to days. Partners on the broader effort include researchers at MIT, the University of Cambridge, the University of Edinburgh and Calico Life Sciences, reflecting an attempt to test the approach across very different fields.
The honest framing matters. Co-Scientist proposes hypotheses; it does not prove them, and every lead still demands laboratory work, peer review and, for any drug, years of clinical testing before it could reach patients. Multi-agent systems can also generate plausible-sounding but wrong ideas, so human scientists remain firmly in charge of deciding what to pursue. Used carefully, though, a tool that helps overworked labs spot promising directions they might have missed is a genuinely useful addition to the scientific toolkit, one that aims to amplify human curiosity rather than replace it.
