A Google AI reached in days an explanation a London lab had spent years proving
Asked how certain bacterial genes spread, Google's AI proposed the answer an Imperial College team had proved but not yet published.
In February 2025 Google Research announced the AI co-scientist, a group of AI agents built on Gemini 2.0 that work together to propose and rank research ideas.
Its most striking result was about drug-resistant bacteria. Bacteria swap genes with each other, which is how resistance spreads, and the question was how one kind of mobile gene package travels so widely. The system proposed that these packages hitch a ride on the tails of many different viruses that infect bacteria, which lets them reach more species. José Penadés’s team at Imperial College London had already shown this in the lab but had not published it. Both papers appeared in Cell on 9 September 2025. Imperial says the AI reached work “that had taken years of work to establish, but doing so in a matter of days”.
The same system was also pointed at existing drugs that might be reused. For acute myeloid leukaemia, experts picked 5 of its 30 top suggestions to test and 3 killed leukaemia cells in a dish. For liver scarring, 2 of 3 tested suggestions reduced it without harming healthy cells.
What it means
AI can now generate the kind of explanation that takes a specialist lab years to pin down, which could shorten the slow guessing stage of research. It did not prove anything: people ran every experiment, and this was one case where the answer was already known inside a lab, so it was not a blind prediction. Nobody has yet published how often the system’s suggestions turn out to be right.
Caveats
- The AI matched a result the lab had already established, rather than making a prediction in advance that was then tested.
- It is a single, self-selected example, and Google authors co-wrote the Cell paper.
- Human labs ran all the experiments.
- The leukaemia and liver results are early lab work reported on Google’s blog, not results in patients.
- Media often say “48 hours”; Imperial says “a matter of days”.
- There is no published tally of how often the system’s suggestions turn out right.