Chalmers' AI scientist turned 60,000 yeast records into nearly 2,000 testable ideas

According to CORDIS, a robot scientist in Sweden worked through a database of about 60,000 public scientific records covering how genes, chemicals and traits interact in brewer's yeast. From that material it produced almost 2,000 testable ideas, each concerning the way various nutrients influence cell growth and the cell's response to stress. The robot also designed the experiments, gave instructions to other robots, went through the results and revised or adjusted any hypotheses that were disproven.

Chalmers researchers presented the system, called Eve, in the Journal of the Royal Society Interface. Eve handled all of the planning for the yeast experiments by itself and issued the instructions that other robots carried out. Human staff did only simple jobs, such as safety checks and moving plates from place to place.

Eve existed before this study. It was built by Ross King, a Chalmers professor in computer science and engineering, as a general-purpose robot scientist aimed specifically at drug discovery. What is new is the reasoning layer. The Chalmers team rebuilt Eve using large language models and automated reasoning. As a result, it needed no human to generate hypotheses, design experiments, interpret results or update its knowledge when new results came in.

For an R&D operations or lab informatics leader, the more useful question is what had to be true about Eve's data and instructions for the handoff to work, because those are the pieces a company lab controls today.

Inside Eve's experiment loop

Once it had produced its hypotheses, Eve assessed them, chose suitable comparison controls and turned each idea into machine-readable, step-by-step instructions the lab could run. When testing ended, Eve went through the data and checked how well its predictions held up. It then automatically revised or adjusted any hypotheses that had been disproven so they were ready for later experiments.

In a news article, Ievgeniia Tiukova, a postdoctoral researcher in Chalmers' Department of Life Sciences, said the data was more than a human could analyse. The AI scientist, she said, was able to spot promising biological questions and propose experiments to test them. It could then assess the results and keep refining what it understood as new evidence came in. She added that the system does more than support decisions, because it actively produces new scientific knowledge.

For an operator, that last distinction is the one that counts. With a decision-support tool, a person still chooses what runs next. In Eve's yeast work, that choice belonged to the system, and that is why the human jobs could shrink to safety and plate handling.

The chain includes the translation from idea to executable protocol, which requires lab procedures expressed in a form a robot can run. It also includes the return path, where results flow back with enough context, including which controls were used, for the system to judge its own predictions.

What the AI was fed

Corporate lab knowledge sits across instruments, notebooks and enterprise systems, and it reaches analytics tools on whatever schedule IT has set.

The refresh figure traces to Trish Meek, whose Industry Perspectives column, presented by Waters, ran in Drug Discovery News on October 15. Most organizations, she writes, opt for a broad data lake. It keeps critical information from every system, yet never the complete record of any single system, and it is updated anywhere from once a day to once a week. Meek argues instead for what she calls a data pond: a smaller, deep store built for one lab domain, such as the analytical lab. It would contain all of that domain's contextualized analytical data in near real time and pass it into the corporate lake through a streamlined pipeline.

She starts from the question executives keep asking: can we use AI to query our data and get an answer? Of course, she writes. The bigger issue is whether the underlying data is right enough to produce the correct answer.

The pond idea isn't hers originally. Meek credits Prajesh Kumar, who used the term for a "series of context-centric, analytics-driven, purpose-built" ponds that enable focused data science. Her case for it is practical. Building one enterprise lake that meets every need is incredibly difficult, she says, and even a company that managed it would find the pipelines that keep it current difficult and costly to maintain.

Set beside Eve, the point sharpens. A system that revises its hypotheses after every round can only move as fast as its results arrive.

In that sense, Eve's knowledge base worked like a pond. It went deep on one organism and one class of question, the gene, chemical and trait interactions in brewer's yeast, instead of skimming across everything. With a weekly refresh, an Eve-style loop in a corporate lab could revise its hypotheses no more than once a week.

Who notices the lag first

Meek points to QA/QC labs as the sharpest case: they need to analyze their own data without waiting on a corporate data science team, so they can decide before samples degrade or before it's too late to save a production batch. Underneath that is a wider problem. Scientists' questions are specific, she writes. They often need information a corporate data lake would not normally hold, plus a grasp of chemistry, biology or physics that data scientists may not have. Scientists don't need all of an organization's data, only the relevant analytical data.

She also notes that ELN and LIMS vendors are building agentic frameworks into their platforms to simplify laboratory execution while maintaining regulatory and process control. Those systems hold the context that makes a result meaningful, such as whether a batch passed or a candidate looks more promising than its peers. Meek's proposal is to connect them to an analytical data pond, with agents that can query both.

Waters, which presented the column, says it is putting data access first in its informatics products, meaning fewer bottlenecks between data creation and use and better interoperability across instruments, applications and enterprise systems. Even read as a vendor's argument, the setup lines up with Eve. The enterprise system knows what a result means, the pond holds the measurement, and an Eve-style agent would need both to judge its own predictions against its controls.

The practical work Eve's study points to is getting protocols into machine-executable form and getting results, with their context, back to the reasoning layer quickly.

Where yeast makes it easier

Talus Bioscience is working on a different problem. Genetic Engineering and Biotechnology News reported that Talus has released Ptarmigan-1. The company calls it the first structure-free AI model of its kind and says it can accurately predict where small molecules bind on proteins throughout the full human proteome. The model targets proteins that have no stable, defined 3D structure, a group that makes up approximately 40 percent of human proteins in cells. Once outside cells, these proteins fall apart, and even advanced AI tools like AlphaFold produce structures for them that cannot be interpreted.

So Talus measures them where they live. Its MARMOT platform, short for Measuring Modulation of Transcription, is an assay that captures snapshots of proteins inside the cell without removing or tagging them. Going label-free is a deliberate choice: co-founder and CEO Alex Federation told GEN that adding tags changes these proteins' dynamics. "We use these new label-free proteomics methods as our camera to take those snapshots," he said. "That's the fundamental measurement that we're taking."

Federation and co-founder Lindsay Pino, the chief technology officer who leads development of the screening platform, met in Seattle about 10 years ago and have worked together since on the technology underneath it. The goal is concrete. Talus' most advanced molecule targets brachyury, a protein encoded by the TBXT gene that is normally expressed only in embryonic stem cells and that the rare bone cancer chordoma reactivates.

The comparison indicates where an Eve-style loop applies most directly: problems where dense, structured knowledge already exists, publicly or inside the company. Where it doesn't, the knowledge base has to be measured into existence first, and Talus' timeline suggests that can take years.

What stays with the scientists

King is direct about the benefit. In his view, AI scientists could cut the time it takes to work through complex scientific questions and help labs make better use of their resources. He also leaves people in charge of the parts no database can provide. Human scientists remain essential, he says, for setting research priorities, judging broader significance and providing ethical oversight. He expects AI scientists to work alongside them in biology, medicine and biotechnology.

A February 2026 paper in the Hastings Center Report by Mohammad Hosseini, Maya Murad and David B. Resnik lines up closely with Eve's loop. The authors note that scientists have begun using AI agents to review the literature, formulate hypotheses, run virtual tests on them and conduct experiments, and they expect those agents to raise productivity and efficiency. The risks they list include responsibility gaps, deskilling of researchers, loss of entry-level research jobs and AI-generated knowledge that humans cannot verify or comprehend.

Their remedy is institutional. Train researchers in AI and algorithmic literacy, bias identification and output verification, and think hard, soon, about which research tasks should and should not be automated. In Eve's loop, output verification would have to sit at the step where the system scored its own predictions and rewrote the disproven ones. That suggests the human role in an Eve-style lab moves from moving plates to checking what the loop concluded.

For an R&D operations leader, that becomes a design decision to make before any pilot: which steps in the loop get a human check, and who on the team is trained to make it.