Nature Publishes Twin AI 'Research Assistant' Systems That Design and Interpret Real Experiments

Two AI systems that can assist across the full arc of scientific research — generating hypotheses, designing experiments and analysing data — were published in Nature this week, marking the clearest peer-reviewed evidence yet that language-model agents can produce ideas worth testing at the bench.

The systems, developed independently, are Co-Scientist from Google DeepMind and Robin from FutureHouse (now Edison Scientific). Both debuted as preprints in early 2025; their graduation to Nature papers comes with expanded experimental validation.

What They Found

Co-Scientist, a general-purpose multi-agent system built on Gemini and driven by natural-language prompts, proposed new drug candidates and combination therapies for acute myeloid leukemia, discovered drug targets for liver fibrosis, and uncovered genetic mechanisms behind antimicrobial resistance — including independently re-deriving an unpublished hypothesis about bacterial gene transfer that its authors had spent a decade establishing.

Robin, which orchestrates multiple models including OpenAI and Anthropic systems, facilitated identification of a potential treatment for dry age-related macular degeneration — pinpointing a modifiable process within retinal cells and proposing a drug candidate never previously suggested for the condition. Wet-lab experiments confirmed the mechanism.

Assistants, Not Authors

Both teams are explicit that the systems assist rather than replace researchers: humans selected which hypotheses to test, ran the experiments and interpreted edge cases. The papers' significance lies in the pipeline — multi-agent generation, internal critique and ranking, literature grounding — producing hypotheses that survived contact with biology at above-chance rates.

Critics quoted in accompanying commentary warn of a subtler failure mode: overreliance on AI-generated hypotheses could create a closed loop that recycles the literature's existing biases instead of producing genuine novelty. Both systems mine published knowledge; neither observes the world.

The Race Context

The publications land in a crowded season for AI-for-science: Anthropic launched its Claude Science workbench with a drug-discovery program for neglected diseases, Sakana's AI Scientist passed peer review, and benchmark efforts like NatureBench are trying to quantify agentic research capability. What Nature's imprimatur adds is epistemic: the question is no longer whether AI research agents work in demos, but how laboratories restructure around hypothesis engines that cost pennies per idea — and how peer review copes when the ideas arrive faster than the benches that test them.