BacterAI, a new artificial intelligence (AI) created by researchers at the University of Illinois at Urbana-Champaign and the University of Michigan, carried out as many as 10,000 science experiments in a single day without human intervention as it sought to define the metabolic processes of two bacteria.

The AI performed the experiments digitally and then analyzed and categorized the results, eliminating the need for test tubes, Bunsen burners, and other lab gear.

Even more impressive, the scientists gave their AI no information to start with.

Instead, they relied on a technique known as “deep reinforced learning (DRL).”

In DRL, an AI is given a question and then it mines relevant on-line literature for information. Next, it tries random combinations of elements—in this case, various combinations of the 20 amino acids that the two bacteria are made of. Possible combinations numbered more than a million.

Instead of humans spoon-feeding data sets into BacterAI, as conventional methods would have, the AI created its own data sets through a series of games, learning rules of each “game” as it played.

“Learning from a blank slate avoids biasing results with prior knowledge,” the research team wrote in the journal Nature Microbiology.

Once it had the essential knowledge and a set of rules, the AI proceeded to simulate hundreds of combinations of amino acids daily—as many as 10,000—using the previous day’s results to predict which new experiments were most likely to give useful results. 

Instead of simply testing every possible one of the million-plus combinations of amino acids, “BacterAI must select the most informative experiments and train a computational model to predict the results for untested combinations,” the research team explained.

After nine days, 90 percent of the AI’s suggested possible combinations were reasonable, the team reported.  

TREND FORECAST: BacterAI’s method of “blank slate learning” can be turned loose to explore crannies of science that humans would not or could not spend their time and other resources to explore.

This approach will be quickly adopted and adapted for use in a range of other fields, yielding knowledge that might have remained hidden for decades if left to humans alone.