Paul Jensen, Associate Professor of Biomedical Engineering, College of Engineering
Benjamin David, Assistant Research Scientist, Biomedical Engineering, College of Engineering and Michigan Medical School
The Jensen lab’s AI journey began with a borrowed idea: if reinforcement learning can master complex board games such as chess and Go without any prior knowledge, could it learn microbiology?
Game-playing AI systems operate in a loop: design an experiment (a move on a board), receive a signal (win, lose, draw), update an internal predictive model, and design the next experiment. The Jensen lab wanted to use the same AI-driven experiments to study bacteria, so they began developing a platform called BacterAI. What they discovered was that building the AI system was not the difficult problem; the hard part was connecting it to the real world.
In a game, an experiment takes milliseconds. The AI agent makes a move, and a perfect computer simulation of the game responds instantly. In contrast, BacterAI doesn’t know that when it requests an experiment, it is kicking off an entire process in the lab. The agent “just thinks it’s playing a game,” as Benjamin David put it, “and that we’re taking a very long time to make its moves in the lab.” Building the infrastructure to close that loop — using laboratory automation at scale — became the defining engineering challenge of the project, and it took years.
The automation platform the team built is a cascade of custom software and robotics. A scheduling layer assigns BacterAI’s experimental requests to assay plates, randomizing positions and embedding quality control samples. A planning system called Pourfecto translates the plate designs into protocols for a fleet of liquid-handling robots that prepare bacterial cultures. Robotic arms move the cultures between incubators and plate readers on timed cycles, and a data pipeline transforms the resulting growth profiles into the reward signal that trains BacterAI’s internal model. When the team started in 2020, they could run 300 experiments per day. After a year of operational refinement, they were running 10,000.
The scientific focus has been the oral microbiome, asking the most basic question in microbiology: what nutrients does a bacterium need to grow? After over a million experiments, the lab is now expanding the scope dramatically. In partnership with the Align Foundation, they are building what they describe as the world’s largest open dataset for microbial phenotyping — one thousand microbes grown in one thousand different conditions each. The goal is a model that combines a microbe’s genome with its growth environment to predict whether it will thrive: a resource for a field that, as David acknowledged, is “much more data-limited than many other fields.” The dataset is designed to be open, and the engineering modular. Any robotic upgrade or new model architecture is designed to fit into the existing platform rather than require reconstruction from scratch. What began as a proof of concept for a single organism has become infrastructure for a community.