Researchers at the University of Illinois at Urbana-Champaign, and University of Michigan (U-M) have developed an artificial intelligence (AI) system, dubbed BacterAI, which is making it possible for robots to conduct autonomous scientific experiments—as many as 10,000 per day—potentially driving a drastic leap forward in the pace of discovery in areas from medicine to agriculture to environmental science.
Reporting on the “blank slate learning” platform, in Nature Microbiology, the researchers, headed by Paul Jensen, PhD, who is now at the University of Michigan, described how BacterAI mapped the metabolism of two microbes associated with oral health, even though it had no baseline starting information. In their paper titled “BacterAI maps microbial metabolism without prior knowledge,” the investigators explained, ““BacterAI learns by converting scientific questions into simple games that it plays with laboratory robots. The agent then distills its findings into logical rules that can be interpreted by human scientists.”
What Jensen and colleagues term “the microbiome revolution” has identified thousands of species of bacteria that deserve scientific investigation, but, as the authors noted, most species of bacteria will remain unstudied. Artificial intelligence and automation could offer a way of carrying out that study, they continued. “… by replacing humans with algorithms that mine the scientific literature and design new experiments.” However, while the least-studied bacteria would benefit the most from automated research, ironically, the investigators pointed out, this “lack of data makes it difficult to deploy autonomous agents to study these species.”
Deep reinforced learning (RL) is a branch of AI in which agents can solve some games by trial and error, even if they don’t have any prior strategic knowledge, or even know the rules of the game. “Converting biological research questions into games could therefore allow the study of microbes using RL techniques,” the scientists suggested. “We developed an RL agent that solves combinatorially large research questions by ‘playing’ science with automated experiments.”
Bacteria consume some combination of the 20 amino acids needed to support life, but each species requires specific nutrients to grow. The U-M team wanted to know what amino acids are needed by the beneficial microbes in our mouths so they can promote their growth.
“We know almost nothing about most of the bacteria that influence our health. Understanding how bacteria grow is the first step toward reengineering our microbiome,” said Paul Jensen, a U-M assistant professor of biomedical engineering who was at the University of Illinois when the project started.
Figuring out the combination of amino acids that bacteria like is tricky, however. Those 20 amino acids yield more than a million possible combinations, just based on whether each amino acid is present or not. Yet BacterAI was able to discover the amino acid requirements for the growth of the oral bacteria Streptococcus gordonii and Streptococcus sanguinis.
Unlike conventional approaches that feed labeled data sets into a machine-learning model, BacterAI creates its own data set through a series of experiments. To find the right formula for each species, BacterAI tested hundreds of combinations of amino acids per day, honing its focus and changing combinations each morning, based on the previous day’s results. “BacterAI cannot rely on a brute force search of every combination,” the team continued. “Instead, it must select the most informative experiments and train a computational model to predict the results for untested combinations.”
By analyzing the results of previous trials, it comes up with predictions of what new experiments might give it the most information. As a result, BacterAI figured out most of the rules for feeding bacteria with fewer than 4,000 experiments. Within nine days, it was producing accurate predictions 90% of the time.
Jensen added, “When a child learns to walk, they don’t just watch adults walk and then say ‘Ok, I got it,’ stand up, and start walking. They fumble around and do some trial and error first. We wanted our AI agent to take steps and fall down, to come up with its own ideas and make mistakes. Every day, it gets a little better, a little smarter.” Using this approach, the AI system was able to tease out the differences in requirements between the two microorganisms, the team noted, “Learning from a blank slate avoids biasing results with prior knowledge. Using BacterAI, we learned that another oral microbe, Streptococcus sanguinis, has different amino acid auxotrophies from S. gordonii even though the two species are closely related and live in the same environment.”
Little to no research has been conducted on roughly 90% of bacteria, and the amount of time and resources needed to learn even basic scientific information about them using conventional methods is daunting. Automated experimentation can drastically speed up these discoveries. The team ran up to 10,000 experiments in a single day.
“Blank slate learning avoids the need for any prior knowledge of the organism,” the investigators concluded in their paper. “Previous automated biology systems developed and tested hypotheses gleaned from the scientific literature. These projects necessarily studied model organisms with extensive prior knowledge for training the agent’s models. BacterAI’s ability to learn solely from its own data enables the study of the unknown corners of microbiology.”
But the applications go beyond microbiology. Researchers in any field can set up questions as puzzles for AI to solve through this kind of trial and error. “With the recent explosion of mainstream AI over the last several months, many people are uncertain about what it will bring in the future, both positive and negative,” said Adam Dama, PhD, a former engineer in the Jensen Lab and lead author of the study. “But to me, it’s very clear that focused applications of AI like our project will accelerate everyday research.”