Co-Scientist and Robin AI agents deliver validated leukaemia and eye disease drug candidates in Nature studies

/ in E-News

Two new systems, Co-Scientist and Robin, use teams of AI agents to search literature, generate testable hypotheses and even analyse laboratory data. Early results, published in Nature, include drug repurposing candidates validated in cellular assays for leukaemia and eye disease.

Two papers published in Nature this year describe a new approach to biomedical research: putting teams of AI agents to work generating and testing scientific hypotheses, rather than simply retrieving or summarising information. The results suggest these systems can propose plausible, testable ideas that hold up, at least in preliminary laboratory experiments.

Two systems, one shared idea

The first system, Co-Scientist, comes from a large team at Google, led by Juraj Gottweis and colleagues, and was published in Nature on 19 May 2026. It coordinates several specialised AI agents, built on Google’s Gemini models, that generate, critique, rank and refine hypotheses through a tournament-style process modelled loosely on peer debate. The second, called Robin, was developed by Ali Essam Ghareeb and colleagues at FutureHouse, also published in Nature on 19 May 2026. Robin pairs literature-search agents, named Crow and Falcon, with a data-analysis agent called Finch, which can interpret laboratory results such as flow cytometry and RNA sequencing and propose follow-up experiments. Writing in an accompanying News & Views article, Olivier Elemento of Weill Cornell Medicine describes the two systems as evidence of a shift towards “a laboratory discovery cycle with AI involved in every step”. He notes that both draw on earlier proof-of-concept work in autonomous chemistry and protein design, but go further by closing the loop between hypothesis generation and experimental data analysis.

From leukaemia to macular degeneration

Co-Scientist was tested across three biomedical applications of varying difficulty: drug repurposing for acute myeloid leukaemia (AML), novel target discovery for liver fibrosis, and explaining the mechanism behind antimicrobial resistance. Asked to identify novel drug repurposing candidates for AML without any additional human input, the system nominated KIRA6, an inhibitor of the stress-response enzyme IRE1α that had not previously been tested in this cancer. In cellular assays, the compound proved selective for leukaemia cells, killing them at concentrations up to 18 times lower than those needed to harm healthy control cells. Robin, meanwhile, was set the task of finding new treatments for dry age-related macular degeneration, a leading cause of irreversible blindness with few existing therapeutic options. The system proposed boosting a cellular clean-up process called phagocytosis in retinal pigment epithelium cells, and identified ripasudil, a Rho kinase inhibitor already used for glaucoma in Japan, as a strong candidate. According to the authors, ripasudil “increased RPE cell phagocytosis 1.89-fold compared to DMSO controls” in follow-up testing, outperforming an earlier candidate. Robin then suggested RNA sequencing to probe the mechanism, which revealed a threefold rise in expression of ABCA1, a lipid transport gene already linked to macular degeneration susceptibility.

Speed, but also limits

Both papers make a point of quantifying the time saved. The Robin authors report that their system “reduces the total cognitive labour of a discovery cycle from an estimated 872–937 human hours to less than two hours”, having synthesised hundreds of papers in around 30 minutes. Co-Scientist’s authors describe similarly favourable comparisons against human experts and other frontier language models in blinded evaluations of novelty and potential impact. Neither team claims their system replaces the scientist. Robin’s developers are careful to note that their tool “represents a powerful new paradigm for AI-driven scientific discovery” while stressing that its outputs remain hypotheses requiring standard pre-clinical and clinical validation. Similarly, the Co-Scientist authors conclude that their work demonstrates “the promise of meaningfully accelerating scientists’ endeavours”, rather than supplanting them, and both papers describe safeguards such as candidate prioritisation based on established safety profiles and screening for known toxicities. Elemento’s commentary strikes a similarly measured note. He points out that the validated hypotheses in both papers largely recombine facts already present in the literature, such as known drug mechanisms or established biological pathways, rather than originating entirely new biology. Whether these systems can move beyond recombination to genuinely novel discovery, he suggests, remains an open question.

What this means for the laboratory bench

For clinical laboratory scientists, the more immediate relevance may lie in the data-analysis side of these systems. Robin’s Finch agent independently analysed flow cytometry and RNA sequencing datasets, arriving at conclusions that were subsequently confirmed by human analysts working on the same raw data. Both papers also flag the need for firmer connections to automated laboratory platforms, better access to negative and unpublished results, and improved fact-checking against experimental databases, as priorities for the next phase of development. Both groups have also opened limited external access to their systems, with FutureHouse’s Robin components openly available, while Google’s Co-Scientist is being made accessible to selected research groups.

References
Gottweis, J., Weng, W.-H., Daryin, A., et al. (2026). Accelerating scientific discovery with Co-Scientist. Nature. https://doi.org/10.1038/s41586-026-10644-y

Ghareeb, A. E., Chang, B., Mitchener, L., et al. (2026). A multi-agent system for automating scientific discovery. Nature. https://doi.org/10.1038/s41586-026-10652-y

Elemento, O. (2026). AI devises hypotheses and ways to test them. Nature. https://doi.org/10.1038/d41586-026-01873-2