Two AI Systems Use Agent Teams to Generate Hypotheses and Analyze Lab Data

Two new AI systems that coordinate teams of specialized agents to generate scientific hypotheses, design experiments, and interpret data were described today in Nature. One system, developed by researchers at Google DeepMind, is called Co-Scientist. The other, created by the non-profit FutureHouse laboratory in San Francisco, is named Robin.

In one test, Co-Scientist reviewed approved medicines for possible use against acute myeloid leukaemia. Human researchers chose five candidates from the list the system produced; three showed activity in cell-culture experiments. Robin was given the task of identifying drugs for dry age-related macular degeneration.

It first directed literature-review agents to summarize existing studies, then selected experiments for human teams to run. The resulting data were returned to a data-analysis agent for further processing.

Both systems still require human input at multiple stages, including experiment selection and physical laboratory work. Researchers reported that the AI-assisted process produced candidate lists within hours. ” He added that the goal is to give scientists additional analytical capacity. The two papers appear in the current issue of Nature.