Google Gemini-based AI proposed in 48 hours a hypothesis that took scientists 10 years

An AI system based on Google Gemini proposed in two days a hypothesis that required scientists about ten years of laboratory experiments. The subject is the mechanism by which mobile genetic elements are transferred between bacteria: the most likely version generated by the model fully matched the result of years of research, which had not yet been published at the time of the test.

What are cf-PICIs and what was the problem

The subject of study was mobile DNA fragments called cf-PICIs, capable of transferring antibiotic resistance genes between different bacterial species. For a long time, researchers could not explain how these elements penetrate new cells, since they have no delivery mechanism of their own. Years of experiments ultimately showed that cf-PICIs form a protective protein shell around themselves and, to penetrate, effectively "borrow" the tail structures of other viruses — bacteriophages. It is precisely this foreign delivery system that allows them to transfer genetic material into bacterial species they previously could not reach.

How the AI reproduced a ten-year conclusion in 48 hours

To test the model's capabilities, the scientists provided the system with a brief description of the problem, without revealing their own results. The key point is that at that time the answer had not yet been published, so the AI could not simply find a ready-made explanation in open sources. In 48 hours, the system generated five possible scenarios, and the most likely hypothesis accurately reproduced the experimentally discovered mechanism of hijacking viral tail structures.

Leak check and Google's response

The answer was so detailed that Professor Penades contacted Google to make sure that his unpublished data had not accidentally leaked online. The company confirmed complete integrity and the absence of any leak, ruling out the version that the model simply reproduced a hidden fragment of a future publication.

New directions and the limitations of the technology

The other four hypotheses proposed by the AI pointed to additional research directions that biologists had previously not considered and which are now being tested experimentally. At the same time, the authors of the study emphasize: the model did not make an independent discovery and is not capable of replacing real work with biological samples. Such algorithms function as an ultra-fast filter, allowing scientists to instantly rule out ineffective versions and focus on promising hypotheses even before the start of costly laboratory trials.