Machine learning takes the guesswork out of catalyst creation

The traditional approach to developing new catalysts for chemical reactions is basically to ‘try it and see’, by conducting numerous repeated experiments with potential candidate molecules. The good news is, machine learning can make this task much more efficient by predicting the performance of catalysts ahead of time based on theoretical characteristics. Now, researchers from Osaka University have used a computer library of molecules that have been synthesised together with molecules that are entirely theoretical at the moment to find the best catalyst for a specific chemical reaction.

The objective of the team’s work was to find better ways to add groups of carbon to amino acids and peptides, which are very common in living organisms, to modify the properties of these compounds. Like many reactions, these processes are enhanced by catalysts, but a traditional metal-based catalyst is often toxic and/or expensive.

The researchers aimed to use triarylboranes as catalysts, but because of their relatively complex structures, there are potentially hundreds of possibilities. These compounds are based on boron, which is a main group element that is relatively inexpensive and less toxic.

“The assessment of molecular catalysts for organic synthesis can be extremely time-consuming,” said lead author Yusei Hisata. “In the case of the triarylboranes used in our work, many permutations of molecular structures could require months of study just to identify the optimum candidate.”

The researchers combined experimental data from a limited number of synthesised triarylboranes with properties predicted for other molecules that have not yet been synthesised, using theoretical calculations, to make a library of 54 possible catalysts.

“This process assessed parameters that we predicted would affect the reaction progress,” said corresponding author Yoichi Hoshimoto. “These included factors such as the molecular orbital energy levels and the energy barriers to certain processes.”

A Gaussian process regression using the in-silico library identified a promising candidate, and tests with this triarylborane demonstrated a high level of performance. This compound could promote the reactions of an amino acid in very high yields and tolerate the presence of numerous different functional groups. As an added benefit, these reactions generate only water as a harmless coproduct because they successfully used molecular hydrogen, H₂, as a reagent. The study also examined other ways to lower the environmental impact of the process and found that the hazardous solvent tetrahydrofuran could be replaced with the less toxic 4-methyltetrahydropyran.

As modern-day chemists juggle developing new syntheses with limited peers while considering environmental impact, efficiency, cost, sustainability and other factors, the new study demonstrates an important step forward in the use of machine learning to streamline the development of new chemical processes and highlights how these new processes can incorporate changes that work together to generate green systems. The results have been published in the journal Nature Communications.

Image credit: iStock.com/poba