DeepMind's Co-Scientist Publishes in Nature: A Multi-Agent System with Scientific Hypotheses that Researchers Are Validating

DeepMind's Co-Scientist Publishes in Nature: A Multi-Agent System with Scientific Hypotheses that Researchers Are Validating

When researchers at Calico Life Sciences fed Google DeepMind's Co-Scientist system a question about the biology of aging, they weren't expecting a hypothesis they'd later confirm experimentally. But that's what happened. The AI system proposed a novel hypothesis about the integrated stress response, and Calico's scientists subsequently validated it in lab tests. It's a single data point, but it represents something that has been genuinely difficult to demonstrate: an AI system proposing a testable scientific idea that turns out to be correct.

That result, along with several others from collaborators at Stanford, MIT, Cambridge, and Edinburgh, forms the core of a Nature paper published by DeepMind on May 19, 2026. The paper introduces Co-Scientist, a multi-agent system built on Gemini that is designed to do one specific thing: generate, critique, and refine scientific hypotheses at a scale and speed that individual researchers cannot match alone.

Seven Agents and a Tournament

The architecture of Co-Scientist is worth understanding because it explains why the outputs look different from what a standard large language model produces. The system runs a coalition of seven specialized agents, organized around three phases.

In the first phase, a generation agent proposes initial hypotheses grounded in scientific literature and databases, while a proximity agent maps those hypotheses to ensure the system is exploring a diverse range of research directions rather than clustering around obvious territory. In the second phase, a reflection agent acts as a virtual peer reviewer, and a ranking agent orchestrates what DeepMind calls a "tournament of ideas," using pairwise comparisons and simulated scientific debates to identify which hypotheses are most worth pursuing. In the third phase, an evolution agent refines and combines the top-ranked hypotheses, while a meta-review agent synthesizes insights from the debates to continuously improve the system's output and produce a final research proposal.

Coordinating all of this is a supervisor agent that breaks down high-level research goals and runs agents in parallel across multiple directions simultaneously.

The tournament mechanism draws from principles used in AlphaGo and AlphaStar, using Elo-based scoring to rank hypotheses as agents debate them. The critical design choice, according to DeepMind, is where the computation goes: the majority of the system's processing is dedicated to verifying hypotheses, not generating them. The system cross-checks claims against scientific literature, web search, and specialized databases including ChEMBL for chemical data and UniProt for protein information. In select research collaborations, the system can also call on AlphaFold as a tool.

What the Lab Results Actually Show

The most direct test of any hypothesis-generation system is whether the hypotheses hold up experimentally. Co-Scientist has a limited but meaningful set of results in that category.

At Stanford University School of Medicine, Professor Gary Peltz was searching for drug-repurposing candidates for liver fibrosis, a chronic scarring condition with limited treatment options. Co-Scientist surfaced overlooked candidates from existing approved drugs. One of those candidates, when tested in the lab, blocked 91% of a scarring-linked cellular response. The results were published in Advanced Science. Peltz described the system as feeling "like a collaborator that's read everything available about biomedical science, with the reasoning capabilities to find the connections that we're currently missing."

At Calico Life Sciences, Dr. Matt Onsum and Katherine Labbé were using Co-Scientist to explore the biology of aging. The system generated a hypothesis about the integrated stress response that was later confirmed in their lab. Onsum noted that what surprised him was "how much it thinks like a scientist," working naturally with existing scientific reasoning rather than producing outputs that researchers then had to translate.

At MIT, Associate Professor Ritu Raman used Co-Scientist to digest complex literature on ALS, propose testable ideas, and identify where her lab's expertise could be combined with complementary work. That process connected her with Ryan Flynn's lab for potential RNA-based approaches to the disease. The system didn't run experiments. What it did was compress the time needed to identify productive collaborative directions.

At the University of Edinburgh, Professor Filippo Menolascina used Co-Scientist to work on metabolic liver disease. The system proposed hypotheses about disease mechanisms and drug combinations, and generated an explanation for why an existing drug benefits only some patients, an idea that Menolascina's lab tests later supported. At the University of Cambridge, Professor Clare Bryant used the system to narrow a search for proteins responsible for severe disease when flu and COVID-19 pathogens jump from animals to humans, getting to specific amino acid targets that her lab will now test, potentially cutting years of experimental work to months.

In each of these cases, the researchers are clear about what Co-Scientist did and did not do. It processed literature at a scale humans cannot, identified connections across disparate research threads, and proposed hypotheses that were specific enough to test. It did not run experiments. The scientific judgment about whether to pursue a lead, and the experimental work to test it, remained with the researchers.

Development and Safety Review

The system was developed in collaboration with researchers from more than 100 institutions, and DeepMind conducted independent evaluations for potential misuse in Chemical, Biological, Radiological and Nuclear domains before publication. From those evaluations, the team built custom safety classifiers designed to flag unethical research goals and suppress unsafe information from the system's outputs. An enterprise version of Co-Scientist is already being tested by Daiichi Sankyo, Bayer Crop Science, and the US National Laboratories as part of the Genesis Mission.

The individual researcher version is now becoming available through a tool called Hypothesis Generation, accessible via labs.google/science, where researchers can register interest. The system will also expand to more Google Cloud enterprise partners.

What This Changes, and What It Doesn't

The honest framing here is that Co-Scientist addresses a specific bottleneck: the time it takes a researcher to survey a field, identify gaps, and formulate testable hypotheses. Omar Abudayyeh at MIT described using the system as "having a team of 50 people at your disposal, doing all the work within a day." That framing is useful. The system compresses literature synthesis and hypothesis generation, tasks that currently take researchers months, into hours or days.

What it doesn't do is replace the judgment required to design experiments, interpret results, or navigate the institutional and collaborative structures through which science actually moves. Ritu Raman's comment is instructive: "Science is a team sport. Co-Scientist can't do science by itself, and I can't do it all by itself either." The system serves as an accelerant for one specific part of the research cycle, not a replacement for the cycle itself.

The broader significance of the Co-Scientist results is about what they suggest for the pace of scientific iteration. If hypothesis generation can be compressed by an order of magnitude in time and scope, and if even a small fraction of those hypotheses hold up under experimental scrutiny, the downstream effect on research productivity is substantial. The liver fibrosis result at Stanford, the stress response finding at Calico, the ALS collaboration at MIT, all of these moved faster than they would have through conventional literature review and hypothesis formation.

The publication in Nature, paired with the rollout of Hypothesis Generation, puts a concrete tool in the hands of working researchers for the first time. The early results from laboratory validation are modest in number but clear in direction. Hypotheses generated by an AI system, when rigorously verified against literature and databases before they reach a scientist's desk, appear to be worth testing.