Key Insights
- AI researchers are developing systems for research reasoning tasks, including hypothesis generation, experimental design, and data analysis.
- Robin, a tool from FutureHouse, can generate hypotheses, propose experiments, and analyze their results to conduct drug repurposing for any disease.
- Google DeepMind has announced two new tools: Co-Scientist reviews scientific literature and generates hypotheses, while Empirical Research Assistance writes software for researchers.
Two artificial intelligence companies have published studies of agent-based AI tools for research reasoning today. Computer scientists at Google DeepMind and FutureHouse say that two tools, dubbed Co-Scientist and Robin, can propose research hypotheses based on existing scientific literature (Nature 2026, DOIs: 10.1038/s41586-026-10644-y; 10.1038/s41586-026-10652-y). A third tool from DeepMind, Empirical Research Assistance, creates scientific software for researchers (Nature 2026, DOI: 10.1038/s41586-026-10658-6).
The developers say the goal is to assist human scientists rather than replace them. But the tools have limitations: they’re built only on open-access datasets, and it remains to be seen whether they can produce genuinely novel scientific insights or lead to drugs for hard-to-treat diseases.
One area where both companies showcase their platforms’ abilities is drug repurposing—or using US Food and Drug Administration–approved drugs for new disease indications. Some chemists doubt the value of repurposing, though advocates point out that using molecules already known to be safe can cut the time it takes to test efficacy. Both companies also say their models’ capabilities extend to other areas of basic and applied research.
The new papers show how industry is increasingly adopting AI agents, which are programs that enable algorithms to perform tasks with some level of autonomy. Two of the tools described are multiagent systems, which use a number of agents that work alongside one another.
Robin
At FutureHouse, researchers set out to “make the first AI-generated discovery in an end-to-end system,” says Michaela Hinks, who led the team that built Robin. To do that, she says, “we hooked up multiple different general agents that were good at particular tasks.”
Robin comprises three agents: two that conduct literature review and one that analyzes experimental data. All human scientists have to do is enter the name of a disease, Hinks says. Then, a list of hypotheses for treating it will appear—along with specific experiments to test those hypotheses.
As a proof of principle, FutureHouse scientists attempted a repurposing workflow for dry age-related macular degeneration (dry AMD), a leading cause of blindness. With a query from the researchers, the tool reviewed the dry AMD literature, hypothesized a list of FDA-approved drugs that might help, and designed preclinical assays to determine whether they help normalize cell processes and boost survival in retinal cells. The researchers dutifully carried out the proposed RNA sequencing and flow cytometry experiments and returned the data to Robin, which analyzed the results and proposed follow-ups.
In practice, Robin could propose experiments very rapidly; obtaining reagents and waiting for cells to grow took weeks. “The challenge is that AI moves so much faster than biology,” Hinks says. But the researchers came away with two drugs that they thought might treat dry AMD, one of which they say had not previously been applied in this context.
FutureHouse is a nonprofit cofounded in 2023 by Sam Rodriques, a physicist by training, and Andrew White, a professor of chemical engineering at the University of Rochester. In 2025, the organization spun out Edison Scientific, a for-profit arm, with $70 million in funding. Hinks is now the head of science at Edison.
Co-Scientist
In their paper, Google DeepMind researchers describe the multiagent AI platform Co-Scientist as a “structured scientific thinking engine.”
Like Robin, Co-Scientist includes agents that survey the scientific literature, generate testable hypotheses, review and rank them against one another in a simulated debate, and further refine them with reference to the literature. It returns its best guesses to the human prompter. The system differs from Robin in that it includes more agents and thus far does not include reasoning about experimental data.
To evaluate the software, DeepMind scientists asked it to repurpose existing drugs for acute myeloid leukemia (AML). The DeepMind team asked oncologists to review Co-Scientist’s top drug proposals, then tested a handful of them in the lab; several of the suggested drugs kill AML cells with nanomolar potency. One of the most potent, binimetinib, had been through clinical trials in AML but failed at Phase 2.
A cancer researcher who did not want to be named because of concerns about digital backlash tells C&EN that the work is an interesting early step but that Co-Scientist has not identified any especially novel targets. They added that their excitement was tempered by questions—some of which the authors also noted—such as the potential for toxicity, the rationale for drug combinations, and the reason different cell lines had responded differently to the same treatment.
Scientists also deployed Co-Scientist to identify what DeepMind says are novel drug targets implicated in liver fibrosis and then suggested drug repurposing to target them, in a study published last year (Adv. Sci. 2025, DOI: 10.1002/advs.202508751).
It is a long path from promising lab results to an approved drug. Some studies indicate that 9 out of 10 drug candidates fail after the start of human clinical trials—though advocates say repurposed drugs, which have already proved safe in human trials, may have a better chance.
The authors recognize that Co-Scientist cannot overcome those barriers. “While the results are promising, translating these predictions from Co-Scientist into clinical practice will be highly challenging, as the complexity of a disease model, patient heterogeneity, and disease variability cannot be fully captured in such limited in vitro experiments,” they write.
Drug repurposing proposals remain hypotheses until they are proved in clinical trials—and as with binimetinib’s failure, these can be hard to predict. But the researchers caught Co-Scientist hypothesizing correctly at least once: when a lab at Imperial College London used the program to analyze a complex bacterial genetics problem they had recently solved but had not published (Cell 2025, DOI: 10.1016/j.cell.2025.08.018).
A fast-moving field
AI research moves fast, and scientific publication moves slowly: the papers describing Robin and Co-Scientist were submitted more than a year ago. In that time, Co-Scientist has already been used in at least three published papers.
Citing concerns about safety and intellectual property, DeepMind did not release the code behind Co-Scientist but invites researchers to contact the company to collaborate.
Meanwhile, White, who is also Edison’s cofounder and chief technology officer, says the company has built on Robin’s success to create Kosmos, another cluster of AI agents that works on the same principles but can follow more complex research pathways. This advance makes it suited for “broader areas like material science, basic biology, and for translational medicine,” he says, adding that about 75,000 researchers use Kosmos, which is free for academics.
“Soon, these systems will seamlessly interface with automated wet labs, allowing them to independently propose hypotheses, direct robotic platforms to validate them experimentally, and continuously refine their models based on real-time results.”
DeepMind’s second tool, also published in Nature today, enables a researcher to vibe code a software pipeline. A researcher inputs a problem and some ideas for how to resolve it; the large language model churns out code optimized to hit a quality metric. The researchers tested its ability to generate analyses for two problems that draw on notoriously noisy data: single-cell genomics experiments and epidemiological forecasting.
Harvard Medical School biochemist Omar Abudayyeh says in an email to C&EN that he is “incredibly excited about the broader trajectory of AI agent systems in science. . . . Soon, these systems will seamlessly interface with automated wet labs, allowing them to independently propose hypotheses, direct robotic platforms to validate them experimentally, and continuously refine their models based on real-time results.”
Jonathan Gootenberg, who runs a lab jointly at Harvard with Abudayyeh (both were acknowledged in Google’s Co-Scientist paper), adds that the real test of any AI-driven discovery system will be the insight it produces. “While we are optimistic about the field, the product of these tools will be the final judge.”
