Designer enzymes take a big step forward

While protein design is experiencing a period of rapid development and advancement, there’s one particularly challenging class of proteins that’s giving researchers a hard time: enzymes. Now researchers from David Baker’s group at the University of Washington’s Institute for Protein Design have made a significant step forward. According to results published in Science Feb. 13, they used two artificial intelligence programs to design serine hydrolase enzymes to catalyze ester hydrolysis (2025, DOI: 10.1126/science.adu2454).

Crafting enzymes is more difficult than designing other proteins because it requires not just predicting the structure of a single protein conformation, but of changes in conformations as enzymes catalyze reactions. The placement of amino acids must be extremely precise as well.

The researchers used the existing AI program RFdiffusion and the new AI program PLACER to identify designs for serine hydrolase enzymes to test in the laboratory. RFdiffusion, which generates protein structures, designs protein backbones around an active site, while PLACER winnows the results from RFdiffusion to a number of possibilities that are feasible to test in the real world.

The researchers gave RFdiffusion only general directions about the desired enzymes.

“We know what substrate we want to act on, and then we also know three or four amino acids and their positions in relation to the substrate that we want to see in the final design, because those are the catalytic residues that are going to do the chemistry,” says Anna Lauko, one of the project’s leaders. “And the task is, give us a protein that has these amino acids and has a binding site for that substrate.”

This enzyme design process does have some limitations. Serine hydrolases and ester hydrolysis were chosen as the enzyme class and reaction, respectively, because they are very well characterized. The model has not been verified with other enzyme classes yet. Lauko also says that more-complicated activation sites that include more catalytic residues are harder for RFdiffusion to compute. Additionally, while the model can handle α-helices and β-sheets, loop structures remain challenging, in line with other protein design tasks that don’t involve enzymes.

Peter Kast, a lecturer in the Department of Chemistry and Applied Biosciences at the Swiss Federal Institute of Technology (ETH), Zurich, who was not involved with this research, says that while this pathway is a step forward, it will require more optimization. He noted that even though the designed enzymes were more efficient than those from any previous attempt, they were still less catalytically efficient than a naturally occurring esterase. Additionally, doing this hydrolysis with activated esters should be relatively easy because these activated esters aren’t very stable anyway, he says. “I think protein design to make a very efficient enzyme is really still the holy grail of the field.”