Chemists create matchmaking app for biocatalysis

Enzymes are great at building complex molecules, and in some cases greener and safer than traditional chemical reagents. But they can also be notoriously picky about the molecules they work on, which can make it hard for chemists to incorporate them into organic synthesis. All too often, finding the right biocatalyst for a given molecule means screening a big library of enzymes, which can be “a nonstarter” for many organic chemists, says University of Michigan–Ann Arbor biocatalysis researcher Alison Narayan.

That’s why she and her group teamed up with machine-learning-for-chemistry expert Gabe Gomes and his group at Carnegie Mellon to develop Catnip, a web app that predicts which biocatalysts are likely to work for a given molecule and vice versa.

The researchers published the work in a preprint (ChemRxiv 2024 DOI: 10.26434/chemrxiv-2024-w4dtr). Narayan presented the project Wednesday at the American Chemical Society Spring 2025 meeting in the Division of Organic Chemistry. Carnegie Mellon PhD student Daniil Boiko, who spearheaded the algorithm’s design, had also talked about it on Tuesday in the Division of Computers in Chemistry.

Alexandra Paton, a postdoctoral researcher in Narayan’s lab, took 314 sequentially diverse α-ketoglutarate-dependent nonheme iron enzymes and used high-throughput experimentation to test each one with over 100 different small organic molecules. The researchers chose α-ketoglutarate-dependent enzymes because they are capable of a variety of carbon-hydrogen activations and are easy to scale up and engineer for specific reactions, Narayan says.

Almost a third of the screening combinations were successful, and the team discovered 215 new enzymatic reactions—mostly hydroxylations and eliminations.

Then came the tricky part: using the screening data to build a usable prediction tool. The researchers combined the 215 reactions they had discovered with 139 other known α-ketoglutarate-dependent enzyme reactions from the literature to map the connections between the enzyme space and chemical space they had surveyed. They then used the data to develop a machine learning algorithm.

For any given small organic molecule, Catnip first figures out where it sits in chemical space and which known molecules from the data set most closely resemble it. The model then comes up with the 10 enzymes that it thinks might instigate a successful C-H activation, providing a starting point for chemists to screen or engineer for the reactivity they need. The researchers tested it out with two natural products and a synthetic steroid; in each case they obtained multiple leads for enzymatically modifying the molecules.

Both Narayan and Gomes say they’ve received a lot of positive feedback on the preprint and the app. Even some industry folks are curious about it, Narayan says.

The two groups are continuing to work together to create a similar model for oxidative couplings using cytochrome p450 enzymes—an even more daunting task as the reaction needs two starting molecules rather than one. They also recently added a reverse-search function so that people can plug in an enzyme sequence and see what small molecules it might be compatible with.

Biocatalysis researcher Kyle F. Biegasiewicz of Emory University says the Catnip platform is “a perfect example of how combinations of emerging technologies” in chemistry, biology, and data science have “the potential to entirely change the way we think about doing routine chemical synthesis” with enzymes. “This work is just the beginning of something very special,” he says.