Google DeepMind and Edison Scientific AI Systems Publish Results in Nature

Two rival teams building an "AI scientist" published results on their systems in Nature in May, the world's most prestigious peer-reviewed scientific journal. Google DeepMind's Co-Scientist and Edison Scientific's Robin are multi-agent programs designed to automate successive stages of the scientific method, from generating hypotheses through designing experiments to interpreting data.

What these systems do

Co-Scientist operates as a coalition of specialized agents built on the Gemini model, directed through natural-language queries. The system scales compute at inference time, letting it iteratively refine hypotheses rather than produce a single answer outright. The project is led by a team headed by Demis Hassabis, the 2024 Nobel laureate in chemistry and head of Google DeepMind.

Robin, built by the non-profit FutureHouse (backed in part by former Google CEO Eric Schmidt), runs on OpenAI's o4-mini and Anthropic's Claude 3.7 models. FutureHouse is led by Sam Rodriques, formerly head of a research group at the Francis Crick Institute in London. An updated version of the system, Kosmos, released in November 2025, now searches through 175 million full texts of publications, clinical trial registries and patents.

Concrete lab results

What sets these reports apart from earlier "AI for science" announcements is that the results were verified in the lab. Robin proposed ripasudil, a drug used to treat glaucoma, as a potential therapy for dry age-related macular degeneration, and also flagged the circadian clock modulator KL001 for the same purpose. Both proposals were confirmed in tests on retinal pigment epithelium cells taken from patients.

Co-Scientist, meanwhile, has helped researchers find new uses for existing drugs in acute myeloid leukemia and explain mechanisms of antibiotic resistance. Gary Peltz, a geneticist at Stanford University, is using the system to search for liver fibrosis therapies. Biologists Omar Abudayyeh and Jonathan Gootenberg used Co-Scientist to speed up research into reversing cellular aging, with the system analyzing decades of scientific literature and proposing new genetic targets that rejuvenated cells in lab tests.

From months and years to minutes and hours - Vivek Natarajan, research scientist at Google DeepMind, on the pace at which Co-Scientist is meant to help scientists reach answers.

Why Nature peer review matters

Publishing in Nature is a different thing from a product launch or a company blog post. Both teams went through scientific peer review, in which independent experts scrutinized their methodology and the reliability of their results before publication. For an industry where accusations of overhyped AI promises come up regularly, that kind of validation carries a different weight than ordinary marketing announcements.

AI-assisted discoveries also raise questions about how to divide credit between the system and the human team that designed it and verified the results. In both cases, the systems don't replace the lab, they shorten the stage of generating and screening hypotheses, which is still followed by testing on live cells or animals.

What it means for biotech

The traditional drug discovery process typically takes more than a decade and consumes hundreds of millions of dollars, largely because of the number of dead ends that have to be tested before a working candidate turns up. Systems like Co-Scientist and Robin promise to shorten precisely that early stage, screening scientific literature and databases faster than a human team can.

Edison Scientific also recently announced a partnership with pharmaceutical company Incyte to fold Robin and Kosmos into its in-house drug discovery and development process. That's a sign the tools are moving beyond academia into commercial R&D teams, where speed and the cost of successive clinical trial phases matter above all.

For Polish scientific institutions and biotech companies, it's further evidence that AI research tools are ceasing to be an experiment and becoming part of the standard toolkit. Centers such as the Instytut Biologii Doświadczalnej PAN (the Polish Academy of Sciences' Institute of Experimental Biology) and domestic drug-design firms will need to decide whether and when to bring similar systems into their own research processes, before competitors abroad do.