Eric Schmidt and Suhas Mahesh propose that scientific AI should choose the next experiment instead of waiting for AlphaFold-scale datasets
Eric Schmidt and Suhas Mahesh propose that scientific AI should choose the next experiment instead of waiting for AlphaFold-scale datasets
On August 10, Eric Schmidt and Suhas Mahesh published an article about research agents. They propose using these agents for problems that require scientists to combine disparate data, formulate a hypothesis, and test it in the next experiment.
AlphaFold predicts a protein's three-dimensional structure from its amino acid sequence. The model was trained on the Protein Data Bank, a large archive of experimentally determined structures. Schmidt and Mahesh present AlphaFold as an example of a problem with many comparable measurements and one clearly defined question.
In routine experimental work, biologists and chemists combine computational results, known structures, and several new tests. Cells grown in the laboratory can change their properties, chemical reagents can contain impurities, and experimental conditions vary between laboratories. A new experiment helps researchers distinguish between competing explanations for an observation and decide what to investigate next.
“The ability to do science consists of combining results from different tools and revising conclusions as evidence accumulates,” Schmidt and Mahesh write.
As an existing example, the authors cite Co-Scientist, a multi-agent AI system based on Gemini. In its cellular aging case study, the system proposed genetic factors for testing, some of which the laboratory confirmed experimentally. A scientist gives the system a research objective and provides existing scientific data. Some agents generate candidate hypotheses, while others criticize and refine them. The resulting proposals are then tested in the laboratory. The developers of Co-Scientist evaluated the system on three biomedical tasks: finding new uses for existing drugs, identifying targets (molecules or cellular processes that a drug could act on), and explaining the mechanisms of antibiotic resistance.
Schmidt and Mahesh argue that an agent should take part in this sequence by selecting the next test and recording the data and reasoning behind that choice. The laboratory could then review how the agent arrived at the hypothesis, test it experimentally, and revise the next step in response to the result.