Have AI agents propose research ideas on superbugs
Google's Gemini 2.0-based AI co-scientist, given an Imperial College question on drug-resistant bacteria, reportedly reached the team's unpublished idea in about 48 hours and suggested one new to them.
Done with GeminiGemini 2.0
What they did
Google Research built AI co-scientist, a team of AI agents running on Gemini 2.0. A scientist states a research goal in plain words. A lead agent hands work to helper agents that write ideas, critique them, and rank them in head-to-head contests, using tools like web search. Imperial College researchers asked how resistance to antibiotics spreads. Scientists set the goals and tested chosen ideas in the lab.
How it went
Per the record, the system reached the Imperial team's unpublished explanation in about 48 hours. It also suggested other ideas, one of which was new to the researchers. Google also reports lab confirmation in reusing existing drugs for a type of leukaemia and in proposing new treatment targets.
Worth knowing
The scores the system uses to improve itself come from its own judging, not from an outside check. Experts rated how new and important the ideas were for only 11 research goals.
Try it yourself with Gemini
My research goal is: [research question in plain language]. Search the literature and summarize what is known, with citations. Generate 10 candidate hypotheses that explain [phenomenon]. Critique each one for evidence, novelty and testability. Then compare them head-to-head and rank them, and refine the top three. For each top hypothesis, outline an experimental protocol I could use to test it and say what result would falsify it. Flag any hypothesis that rests on weak or unverified evidence.
Swap the highlighted [parts] for your own details.
Source: Google Research blog · Feb 2025
Written with AI assistance from the original post and checked against our editorial guidelines. Results are as reported by the source, not independently verified.
