Teaching AI to build better enzymes

KEY TAKEAWAYS

• Enzymes can help scientists make valuable chemicals and molecules but engineering them is difficult because they often need to meet several requirements at once. 
• Freeman Hrabowski Scholar Yang Yang’s lab at the University of California, Santa Barbara, co-developed SynFit, an open-source machine learning tool that predicts enzyme variants likely to perform well across multiple traits. 
• Because SynFit can work with small experimental datasets, and its code and data are freely available, the tool could help more labs test new enzymes.  

Enzymes are biology’s tiny chemists. They help build, break, and rearrange molecules quickly and precisely. In the lab, scientists have long put enzymes to work making useful chemicals.  

But nature’s enzymes cannot do every reaction a researcher might want. Sometimes scientists need an enzyme to take on a job it did not evolve to do. That’s where enzyme engineering comes in — researchers make changes to an enzyme, test how those changes affect what the enzyme does, and look for versions that work better.  

The challenge is that a useful enzyme often must get several things right at once. It may need to work quickly, stay stable, produce enough of the desired product, and reliably make that product into the correct shape.  

But improving one feature can cause another to fall apart.  

That makes enzyme engineering a balancing act. “Almost all enzymes have to fulfill multiple requirements,” says Yang Yang, an HHMI Freeman Hrabowski Scholar at the University of California, Santa Barbara.  

Looking Beyond One Trait 

SynFit, an open-source AI tool, was built to help researchers navigate those competing demands. Developed by researchers in Yang’s lab and the lab of his collaborator, Yunan Luo at Georgia Institute of Technology, the tool looks for patterns in experimental data and predicts enzyme variants that are more likely to perform well across several traits at once.  

Traditionally, scientists improved enzymes through directed evolution — a lab process inspired by natural evolution. They make many versions of an enzyme, test them, choose the best performers, and repeat the process.  

Computational models can help narrow that search, but many are designed to predict one trait at a time. So, making an enzyme that has to do several things well can mean screening many variants over multiple rounds. This process can take a lot of time. 

SynFit takes a different approach. Instead of asking which enzyme is best at one trait, the tool learns several measurements at the same time. Researchers start by making and testing a small group of enzyme variants in the lab. Instead of ranking enzymes by a single score, SynFit looks for versions that balance multiple traits at the same time to predict new variants that fit the requirements the researchers care about. 

Testing SynFit in the Lab 

The team tested SynFit to improve a cytochrome c enzyme engineered to help build carbon-boron bonds. These bonds are valuable because they can serve as useful building blocks for making more complex molecules, including those used in medicines and materials. But this particular enzyme-driven reaction is not something nature is already known to do on its own.  

After training SynFit on a small batch of experimental data from the cytochrome c enzyme, the researchers used the model to choose new enzyme versions to test in the lab. Many of those designed variants hit the balance the researchers were hoping for — high yield and high selectivity. 

The team also wanted to know whether SynFit could work beyond this one enzyme. So, they tested it on public protein datasets where the same protein had been measured in more than one way. The tool performed better than models that treated each measurement separately.  

Yang says the team is already applying the model to other enzymes. “The method is not limited to a specific protein, and it’s highly general,” he says. “We are now applying SynFit and other machine learning-guided protein engineering tools to accelerate the enzyme optimization campaign that we are very excited about.” 

Making the Search More Accessible  

SynFit also worked with relatively small amounts of experimental data. Testing enzyme variants can be slow and resource-intensive, especially when researchers need to measure more than one trait. If a model can learn from smaller sets of experiments, it can help researchers spend more of their efforts on those that are most likely to work.  

The team has made SynFit’s code and data publicly available so that researchers can replicate what was done by the Yang lab and try the approach for their own enzyme-engineering projects.  

For enzyme engineers, SynFit could make the search less about testing everything and more about choosing what to test next.  

“We focus on what seems like a niche problem, enzyme catalysis, but these protein language models are opening up possibilities for protein engineering with broad biomedical implications,” says Yang.