Microbiologists from Imperial College London compared the results of a decade of work with the capabilities of artificial intelligence. They tasked Google's Co-scientist system with explaining how mobile genetic elements transfer antibiotic resistance genes between bacteria.
In 48 hours, the AI, built on the Gemini model as a multi-agent system, proposed five ranked hypotheses. The first exactly matched the mechanism that scientists had experimentally confirmed over years of laboratory research.
This concerns cf-PICIs — capsid-forming phage-inducible chromosomal islands. These elements are capable of creating their own protective shells, but to infect new hosts they hijack tail structures from other bacteriophages.
Project leader Professor José R. Penadés from Imperial College London and the Fleming Initiative, together with co-author Tiago Dias da Costa, published the results in the journal Cell. They deliberately chose an already solved but not yet published problem to test the AI against a known answer.
Penadés even wrote to Google asking whether a data leak had occurred. The company confirmed that there had been no access to unpublished results.
The remaining four hypotheses pointed to directions the team had not previously considered. They are now being tested in the laboratory.
The test showed that the system effectively synthesizes complex biological data and proposes viable hypotheses. But can it accelerate discoveries in areas where there is no ready-made benchmark for comparison?
The scientists emphasize: AI acts as a tool for filtering ideas, not a replacement for years of experiments with real samples. The publication in Cell appeared in November 2025.
