Nature papers reveal AI systems aiding drug repurposing for leukemia and macular degeneration

Two papers published in Nature on Tuesday describe the development of artificial intelligence assistants designed to aid scientific hypothesis generation, specifically focusing on drug-retargeting tasks. The research, led by Google and the nonprofit FutureHouse, demonstrates how AI can manage the overwhelming volume of scientific literature to identify non-obvious connections across disparate fields. Both systems operate as agentic tools that function in the background, allowing human experts to review and validate the generated suggestions for further testing.

Google’s Co-Scientist, built on the company’s Gemini large language model, functions as a “scientist in the loop” system. It searches literature and evaluates hypotheses through a tournament process to suggest drug targets for acute myeloid leukemia. The system employs a Reflection agent that accesses external search tools to prevent the hallucination of implausible ideas, while an Evolution agent refines surviving hypotheses. The process prioritises criteria such as plausibility, novelty, testability, and safety, with human experts reviewing the AI’s suggestions before any biological testing occurs.

FutureHouse’s Robin system utilises specialised tools for literature summarisation and data evaluation to propose hypotheses and drug candidates for a form of macular degeneration. Robin analyses 551 papers in 30 minutes, a task estimated to take 540 hours for a human researcher. The system uses specific tools named after birds: Crow for concise summaries, Falcon for deep overviews, and Finch to automate the evaluation of data from standard biological screening assays such as flow cytometry and RNA-seq.

The importance of specialised literature tools was highlighted when FutureHouse tested Robin against OpenAI’s o4-mini model. Swapping Robin’s Crow tool for OpenAI’s model increased the rate of hallucinated references from zero to 45 percent. Furthermore, while OpenAI’s tool suggested drugs that Robin had not identified, those candidates failed to have an effect on the tested cells, underscoring the risks of using generalist models for precise scientific literature retrieval.

Both systems aim to mitigate scientific information overload by identifying connections that human experts may overlook due to the compartmentalisation of knowledge. Google’s system identified drugs effective only against subsets of myeloid leukemia cells, consistent with known biological variability in cancer growth routes. The researchers note that while these successes represent a step forward in repurposing existing drugs, the AI systems are not yet capable of solving the more complex, open-ended problems that characterise much of biological research.