AI-based robot uncovers general reaction conditions for Suzuki-Miyaura coupling

Chemists know that getting conditions right for an individual chemical reaction can be a persnickety process. So figuring out general reaction conditions that work for many different substrates is considered almost impossible. But now, an international team led by Martin Burke and Bartosz Grzybowski has figured out general reaction conditions for the very widely used Suzuki-Miyaura reaction (Science 2022, DOI: 10.1126/science.adc8743). The scientists built a robot that uses artificial intelligence to learn from the outcomes of reactions it tries and then make the reactions work better. Using the system, the team more than doubled the average yield of human-optimized Suzuki-Miyaura coupling reactions. This approach could serve as a blueprint for future automated organic chemistry systems to quickly and easily make a myriad of other types of molecules.

In the Suzuki-Miyaura reaction, chemists snap together organohalides and organoboronates over palladium catalysts to make new carbon-carbon bonds (example shown). But scientists can use a wide range of reagents, solvents, catalysts, and temperatures to make the reaction happen. Finding general conditions that work for most of the possible reagents—crucial for automation of this chemistry—is no small feat. “It’s like this astronomical haystack, and we’re trying to find a needle, right in the middle, which are the most general conditions for . . . most of those couplings,” says Burke, an organic and automation chemist at the University of Illinois Urbana-Champaign.

The team’s answer to this problem links a decision-making tool that predicts the best reaction conditions, a machine learning module that homes in on them, and a robot to carry out experiments using those conditions. The group then analyzed the products with liquid chromatography/mass spectrometry and ultraviolet–visible spectroscopy to determine yields. Overall, the robot explored 528 reactions.

There were two crucial parts to solving this problem, Burke says. First, the team shrank the haystack: instead of looking through every possible combination of reagents, it grouped together similar molecules. Second, the researchers created a closed-loop system, one that could try various reactions, see what didn’t work, and learn from this information, says Grzybowski, a physical organic chemist at the Ulsan Institute of Science and Technology and the Polish Academy of Sciences. The researchers seeded the model with experimental data, and “it sniffed around and built knowledge, then started zooming into the regions where the conditions were better,” Grzybowski says. The robot would do reactions, analyze the results, and do more. “After five rounds, we pretty much knew all the answers, even though we hadn’t sampled the whole space,” Burke says.

The AI often asked the robot to run experiments that it predicted would not work. “It was picking winners and losers at the same time,” Burke says. Once the AI predicted the reaction conditions that would give the highest average yield, the team tested them by running reactions it had not trained the system on. “We didn’t start with the small haystack—we went back out to the whole massive haystack,” Burke says. In this test set of reactions, the robot had an average 46% yield compared with average literature yields of 21%.

The ability to search a large amount of complex research conditions to find general reaction conditions is a significant achievement, says Benji Maruyama, autonomous chemistry expert at the Air Force Research Laboratory in Dayton, Ohio.

The team’s model can potentially be applied to other reactions and unsolved problems in science, Burke says. “There are a lot of interesting problems in chemistry and far beyond, which are these massive search spaces that almost seem intractable. This could be a general playbook to go after solving these very multidimensional problems,” he says.