Google Publishes Co-Scientist in Nature, Opens AI Hypothesis Tool to Researchers

Google has published its Co-Scientist research in Nature, introducing a multi-agent AI system built with Gemini that generates, debates, and evolves novel scientific hypotheses — and it has already produced lab-confirmed results, including a repurposed drug candidate that blocked 91% of a scarring-linked response in liver fibrosis tests.

The company is now moving the system out of the lab. Google announced a new experimental tool, Hypothesis Generation, jointly developed across Google DeepMind, Google Research, Google Cloud and Google Labs. Individual researchers can register interest at labs.google/science, with rollout beginning in the coming weeks. An enterprise-grade version has been previewed with Daiichi Sankyo, Bayer Crop Science, and the US National Laboratories as part of the Genesis Mission.

The stakes are straightforward. Google frames hypothesis generation — the act of connecting disparate facts into a testable idea — as a bottleneck in an era of information overload. If AI can compress that bottleneck, the pace of discovery in drug development, infectious disease, and materials research could shift measurably. The Nature paper and the early experimental record matter because they offer evidence beyond demos.

A coalition of agents, not a single model

Co-Scientist is not one model but a coalition of specialized Gemini-based agents organized into three phases: generate, debate, and evolve.

A Generation agent proposes initial focus areas and hypotheses grounded in scientific literature and data. A Proximity agent maps and clusters those hypotheses to keep exploration of the research space diverse. A Reflection agent acts as what Google calls a "virtual peer reviewer," evaluating hypotheses for correctness, quality, and novelty. A Ranking agent runs an "idea tournament" using pairwise comparisons and simulated scientific debates. An Evolution agent continuously refines and combines the top-ranked hypotheses, and a Meta-review agent synthesizes insights from the debates to generate a final research proposal for the scientist.

A supervisor agent orchestrates the coalition. Google describes it as an adaptive, freeform planner that breaks high-level research goals into executable steps and coordinates agents to run in parallel — a contrast, the company notes, with AI models that think linearly.

The tournament of ideas

To manage the thousands of research directions the system can explore, Google developed what it calls a "tournament of ideas," an Elo-based ranking mechanism that iteratively prioritizes hypotheses while injecting fresh knowledge to expand the search space. The approach draws on principles from AlphaGo and AlphaStar — but the agents play scientific debates instead of games.

The majority of the system's computation goes to verification rather than generation. The system cross-checks claims against scientific literature and data, integrating web search and specialized databases such as ChEMBL and UniProt. It can also use specialized models as tools, including AlphaFold, which Google is testing in select research collaborations.

Validation in the lab

Over the past year, Google collaborated with researchers at more than 100 institutions to test Co-Scientist on complex life sciences problems. Several results stand out.

At Stanford University School of Medicine, Co-Scientist accelerated Gary Peltz's search for liver fibrosis treatments. The system surfaced overlooked drug-repurposing candidates, one of which blocked 91% of a scarring-linked response in lab tests. The results, published in Advanced Science, point toward new gene-regulating approaches to chronic liver disease. "Co-Scientist feels like a collaborator that's read everything available about biomedical science, with the reasoning capabilities to find the connections that we're currently missing," Peltz said.

At MIT, Associate Professor Ritu Raman used the system to digest complex literature on ALS and identify where complementary expertise could strengthen promising leads, sparking a collaboration with Ryan Flynn's lab on potential RNA-based approaches. "Science is a team sport. Co-Scientist can't do science by itself, and I can't do it all by myself either. It helps me structure my thoughts, so I know what to ask of other experts and collaborators," Raman said.

Omar Abudayyeh and Jonathan Gootenberg, principal investigators of the Abudayyeh–Gootenberg Lab, are applying Co-Scientist to research on reversing cellular aging. The system synthesizes decades of literature into novel genetic leads that lab tests have shown can rejuvenate cells, and it cuts the time to analyze large screening datasets from months to days. "Using Co-Scientist feels like having a team of 50 people at your disposal, doing all the work within a day, which isn't something we can otherwise do with our lab," Abudayyeh said.

At the University of Edinburgh, Professor of Engineering Biology Filippo Menolascina used the system to generate hypotheses for metabolic liver disease. Co-Scientist highlighted disease mechanisms and drug combinations and explained why an existing drug benefits only some patients — an idea his lab tests later supported. "Co-Scientist feels like a jetpack for scientists, powering up our ability to identify promising mechanisms. I think we're on the brink of a scientific revolution that will significantly shorten the iteration cycles needed to achieve breakthroughs," Menolascina said.

Clare Bryant, Professor of Innate Immunity at the University of Cambridge, is using Co-Scientist to identify proteins that cause severe disease when pathogens like flu and COVID-19 leap from animals to humans. Working with the system, she narrowed the hunt to specific amino acids her lab will test — potentially cutting years of experimental work down to months. "It catches what I'd miss in a data-rich field and helps me prioritise, so my team can focus on answering the right questions in the lab," Bryant said.

At Calico Life Sciences, Matt Onsum and Katherine Labbé applied Co-Scientist to the biology of aging. The system generated a novel hypothesis about the integrated stress response that was later confirmed in the lab. "What I found both exciting and surprising about using Co-Scientist is how much it thinks like a scientist. It really works naturally with how a scientist already thinks and behaves," said Onsum, Calico's Head of AI/ML.

Safety evaluations and CBRN classifiers

Co-Scientist underwent extensive internal and external safety evaluations, Google said. Because of its proficiency in life and physical sciences, the company also ran independent evaluations for misuse in Chemical, Biological, Radiological and Nuclear (CBRN) domains. Based on those findings, Google built custom safety classifiers to flag unethical research goals and mitigate the surfacing of unsafe information.

The research was led by Juraj Gottweis and Vivek Natarajan, with Alan Karthikesalingam, Annalisa Pawlosky, and Yunhan Xu, and contributions from a long list of Google researchers including Demis Hassabis, Yossi Matias, and Pushmeet Kohli.

Google positions Co-Scientist as a partner, not a replacement for scientific or clinical expertise, and states that users remain responsible for decisions made using its outputs. With individual access rolling out through Gemini for Science and broader Google Cloud enterprise access planned, the near-term question is whether the lab-validated results from Stanford, Calico, and Edinburgh will hold at scale across thousands of new users.