AIBRIEF
Chalmers team publishes a peer reviewed cycle where LLM agents propose a yeast experiment, robots run it, and the loop revises the next hypothesis when the first one fails.
A team at Chalmers University has stitched AI hypothesis-generation, multi-agent experiment planning, robotic execution, and result interpretation into a single inspectable loop inside a yeast-genetics lab, publishing the cycle in the Journal of the Royal Society Interface. In one documented pass, the system's first prediction failed, then the same loop proposed the corrective experiment, the robots ran it, and humans decided what the result meant.
The architecture is narrower than "AI runs a lab." LLM agents guided by symbolic relational learning propose hypotheses into a graph database with controlled vocabularies; automated cell-culture and metabolomics platforms carry them out, per the authors' abstract. Reported yeast observations include glutamate-induced growth inhibition in spermine-treated cells and partial rescue of formic-acid stress by aminoadipate, as ZME Science describes. The platform builds on the earlier Eve robot-scientist infrastructure, adding a published multi-agent planning layer to existing lab hardware.
The first AI prediction in the documented cycle did not hold up, and according to ZME Science, the loop then selected aminoadipate and proposed the next experiment. The team frames the system as moving from a tool scientists use toward something that can carry out chunks of the scientific process itself, with humans setting the agenda, the safety limits, and the interpretation.