AI Ran A Lab Almost Entirely On Its Own, Making New Discoveries

This AI Learned From Its Wrong Predictions and Found a Compound That Partly Shields Yeast

Yeast cells exposed to a compound called spermine became more vulnerable to it after an AI-run lab added the amino acid glutamate. The authors describe that interaction as previously underexplored. That same system later found that a compound called aminoadipate partly protects yeast from formic acid, a chemical stressor. The researchers say this protective effect has not been shown before. Both results emerged from experiments an AI planned and a robotic lab carried out.

Researchers in Sweden and the United Kingdom built the system and described it in the peer-reviewed Journal of the Royal Society Interface. It pairs chatbot-style AI with strict logical rules, then connects everything to liquid-handling robots, automated cell-growing equipment, and a machine that measures chemicals inside cells. Baker’s yeast, the microbe behind bread and beer, served as the test subject.

Humans defined the project’s scope, prepared stock solutions, and checked each final plan for safety. Occasional plate transfers between machines also needed a person, a step the authors say could in principle be eliminated. Beyond that, no person was needed to come up with hypotheses or design experiments.

AI Robot Scientist Turned 60,000 Yeast Facts Into Nearly 2,000 Testable Predictions

Researchers first fed the system a database of roughly 60,000 facts about yeast biology, pulled from public scientific records of how genes, chemicals, and traits relate. A pattern-finding program scanned those facts for consistent links, such as yeast that handles lithium poorly tending to hold less of the amino acid arginine, one of the building blocks of proteins.

Matching those links against published chemical measurements from yeast strains missing a single gene produced 1,933 testable predictions across 16 amino acids. Most guessed how adding a nutrient would change yeast’s ability to survive a stress-causing chemical.

For each amino acid tested, the system picked its top-ranked prediction, plus a comparison amino acid expected to do nothing, much like a placebo in a drug trial. One chatbot agent, built on OpenAI’s GPT-4o, drafted an experiment plan. A second agent chose the best of several drafts, and a third converted that plan into instructions robots could read. The chatbot agents also offered possible biological explanations, though the researchers made little use of them.

Robots then mixed the chemicals and grew the yeast for 20 hours, tracking growth by how cloudy the liquid became. Afterward, machines measured the chemicals inside the cells, and each condition was repeated in about 10 separate yeast cultures. Testing covered five amino acids: glutamate, arginine, proline, glutamine, and lysine.

Two Forecasts Held Up, but Lysine Protected Yeast Instead of Harming It

Arginine delivered the predicted result. By itself, it slightly reduced yeast growth, while caffeine alone cut growth by more than half. Together, the pair lowered growth more than their separate effects combined. A comparison amino acid, alanine, produced a smaller drop, which the authors took as confirmation that arginine was responsible.

Glutamate and spermine also matched the forecast, with a catch. Glutamate made yeast more sensitive to spermine, but glutamate levels inside the cells did not rise. The authors suggest a chemical made from glutamate inside the cell may be doing the real work.

Lysine went the other way. The AI forecast that extra lysine would leave yeast less tolerant of a high-sugar environment, yet lysine strongly protected the cells. Lysine levels inside the cells did rise as expected, so the delivery worked while the biological effect flipped. Other predictions fared worse, since several effects turned out not to be specific to the nutrient being tested.

Glutamate Helped Yeast Under Formic Acid Stress, but a Control Compound Helped Far More

A formic acid test also reversed its forecast. The AI predicted glutamate would lower yeast’s tolerance to the acid. Instead, glutamate delivered a measurable growth benefit. Alanine, the comparison compound, produced a far larger effect that nearly canceled the acid’s toll. A result that shows up with an unrelated compound cannot be credited to glutamate, so the authors labeled the interaction non-specific. Earlier researchers had also reported that glutamate influences yeast tolerance to formic acid.

A Missed Prediction Led the AI Robot Scientist to a Partial Rescue by Aminoadipate

Failed predictions were not discarded. A statistical analysis of chemical data from an earlier experiment ranked substances linked to formic acid tolerance. The top-ranked compound on hand in the lab, aminoadipate, became a fresh prediction, and the system tested it. Aminoadipate partly rescued yeast from formic acid stress, confirming the new guess. Both aminoadipate and formic acid are highly toxic to yeast on their own, the authors caution, so the cells may have hit a ceiling that influenced the result.

Every plan, chatbot exchange, and measurement went into a searchable database, which let the system skip one experiment. A proline hypothesis involving formic acid matched data already collected, since proline had served as a comparison compound earlier. That data showed no drop in tolerance, contradicting the idea, so the test was set aside, saving costly lab resources.

Even so, the system followed a wrong guess to a new compound while its own control compounds flagged shaky results as non-specific, a sign that an AI-run lab can catch some of its own mistakes. Yeast is one organism, though, and whether that holds in messier biology remains untested.

---