DeepMind’s “AI co-scientist” proposes drug ideas that held up in the lab

When an AI system suggests a scientific idea, the obvious question is whether the idea actually works. A peer-reviewed study published in Nature in May 2026 offers a rare and encouraging answer: for Google DeepMind’s “AI co-scientist,” at least some of its proposals held up when researchers tested them at the bench.

The AI co-scientist is a general-purpose, multi-agent system built on Google’s Gemini model. Rather than a single assistant, it orchestrates several specialized agents that generate, debate, rank and refine hypotheses in response to a researcher’s natural-language goal. The Nature paper reports initial validation across three distinct biomedical challenges: repurposing existing drugs for acute myeloid leukemia, discovering new targets for liver fibrosis, and explaining a mechanism behind antimicrobial resistance.

“A peer-reviewed study published in Nature in May 2026 offers a rare and encouraging answer: for Google DeepMind’s “AI co-scientist,” at least some of its proposals held up when researchers tested them at the bench.”

The leukemia work is the most concrete demonstration. The system proposed candidate drugs that could be repurposed against acute myeloid leukemia; oncologists reviewed its top suggestions, and researchers then tested a handful in the laboratory. Several of the proposed drugs inhibited the viability of tumor cells at clinically relevant concentrations across multiple leukemia cell lines — meaning the AI’s hypotheses were not merely plausible on paper but produced real effects in living cells. For antimicrobial resistance, the system independently arrived at an explanation that collaborators had already confirmed experimentally, a striking convergence.

It is important to be precise about what this shows and what it does not. These are early, laboratory-stage results, and a promising signal in cell lines is many careful steps away from a treatment a patient can receive. The DeepMind team frames the co-scientist explicitly as a collaborator: a tool to help scientists generate and prioritize ideas faster, with human researchers always designing the experiments, running them and judging the outcomes.

That framing is where the optimism lies. Modern biomedicine is drowning in literature — no researcher can read even a fraction of what is published in their field — and the bottleneck is often not data but the human time to connect it into testable hypotheses. A system that can survey that vastness and surface a short list of ideas worth testing, some of which then survive real experiments, could meaningfully speed the unglamorous middle of science. Used well, and kept firmly under human judgment, it is a genuinely hopeful tool.