Chalmers researchers test an AI system that proposes and runs experiments

The Chalmers researchers developed a system that generates scientific hypotheses, designs experiments and interprets results, then tested it using brewer’s yeast, Saccharomyces cerevisiae[1]. The setup combines language-model agents with logical scaffolds that incorporate relational learning, structured vocabularies and experimental constraints[2]. Its laboratory framework included robotic liquid handling, automated cell cultivation and sampling, and ion-mobility mass-spectrometry metabolomics[4]. The researchers collected time-series growth data and endpoint metabolic profiles during experiments[3].

Researchers supplied the AI with a Datalog database containing about 60,000 relationships involving yeast phenotypes, physiology and metabolism[3]. The system produced 1,933 independent hypotheses across 16 proteinogenic amino acids[3]. It stored hypotheses, experimental designs and results in a graph database using controlled vocabularies[3]. This structure also retained both results that supported predictions and results that refuted them[3]. For teams assessing the approach, the reported scale describes this yeast study; the evidence does not establish performance across other organisms or research areas[1][3].

Some experiments found measurable effects: a treatment significantly sensitized cells to spermine, at −4.6% per millimolar, and a combined application produced a synergistic decrease relative to independent effects[3]. In a follow-up experiment, aminoadipate partially rescued yeast from formic-acid stress, improving growth by about 7% per millimolar[3]. Not every prediction was confirmed: testing did not support the system’s prediction that arginine would rescue yeast from lithium stress[5]. Together, these findings show both reported experimental effects and a prediction that failed in this test—not a general measure of accuracy across fields[3][5].

The system did not remove people from laboratory work. Human contributions included preparing stock solutions and metabolomics reference samples, as well as occasional transfers of plates and data between machines that were not networked[4]. Researchers also reviewed proposed experiments for safety[4]. Chalmers says human scientists remain essential for setting research priorities, interpreting broader scientific significance and ensuring ethical oversight[1]. The study uses Eve, a system specifically designed for drug discovery[1]. Chalmers says systems of this kind could reduce the time needed to explore complex questions and help optimise laboratory-resource use[1].