An AI system can now take part in a full research loop, from proposing a hypothesis to choosing what to test after an experiment. It still needs people to prepare solutions and move some lab equipment around, which is a useful reminder that “autonomous” often comes with a human loading crew.
From a database to testable predictions
Researchers in Sweden and the United Kingdom built a system to study how amino acids affect the yeast Saccharomyces cerevisiae under chemical stress. Its starting point was a database containing nearly 60,000 facts about yeast, including relationships among genes, chemicals and observable traits.
A pattern-finding algorithm used those facts alongside experimental data to produce 1,933 testable predictions. They covered the effects of 16 amino acids on yeast resistance to different chemical stressors.
Several AI agents then handled parts of the experimental planning. One, based on GPT-4o, drafted an experiment plan. A second chose a suitable plan, and a third turned it into instructions for lab equipment.
Robots tested the ideas
Robots mixed the substances, grew yeast cultures for about 20 hours and measured changes in growth and chemical concentrations inside the cells. Each experimental condition was tested across ten separate cultures. The system stored hypotheses, plans and results in a shared structured database.
Some predictions held up. The system predicted that arginine would strengthen caffeine’s effect. Arginine alone had only a small effect on yeast growth, while caffeine cut growth by more than half. Together, the two had a stronger effect.
Another experiment found that glutamate made yeast more sensitive to spermine. The researchers describe this interaction as insufficiently studied. They suggest the effect may come not directly from glutamate, but from a substance yeast cells produce from it.
A failed prediction still had a use
Not every prediction worked. The system expected lysine to make yeast less tolerant of high sugar concentrations. Instead, lysine noticeably protected the cells. Tests involving formic acid also produced results that were not specific enough to support the predictions, since the control substance had an effect at least as strong as the amino acid being tested.
The researchers did not count these outcomes as confirmations. They did, however, let the system use unsuccessful experiments when choosing what to investigate next. That distinction matters: an automated research system should be able to learn from a result without quietly relabeling it a success.
The database also helped prevent duplicated work. The system could recognize when a proposed experiment had effectively already been run as part of an earlier test, and avoid spending resources on it again.
Autonomy with a human loading crew
The lab was not fully self-running. People prepared the starting solutions, moved some plates between equipment that was not connected, and reviewed experiment plans for safety. But the system could select hypotheses and develop most experimental procedures without human input.
The demonstration was limited to relatively simple hypotheses and one yeast species. It does not show that AI can replace researchers, or that the same approach will work unchanged in more complex biological studies.
What it does show is a shift from AI helping with isolated tasks, such as drafting a protocol or processing results, to taking part in a sequence where one experiment can shape the next. The robots did the lab work, the software decided what to try, and the humans made sure the whole arrangement remained safe and supplied. That is not a scientist in a box. It is a research loop with a few very necessary people standing just outside it.