Researchers use AI to design enzymes that cut a disease-linked protein in cells

Researchers have used AI to design enzymes that cut selected protein sequences, including a disease-associated target inside cells. The work, primarily from David Baker’s University of Washington lab, combines computer-designed structures with laboratory testing and mutations. It demonstrates a route toward custom protein cutters, while exposing a limit: designing the right shape does not guarantee an effective enzyme.

John Timmer described the findings in Ars Technica on October 14. The researchers built two kinds of protease—the name for an enzyme that cuts proteins—using the catalytic machinery of a zinc protease from crayfish and a cysteine protease from yeast.

Keep the cutting machinery, redesign its surroundings

Proteins are chains of amino acids joined by peptide bonds. Those bonds are extremely stable on their own, with an estimated half-life of several hundred years. Proteases accelerate their breakdown, but doing so requires a precise arrangement of amino acids at the enzyme’s active site, where the cutting reaction happens.

The teams preserved the key catalytic amino acids from the existing enzymes, then designed new protein structures around them. The surrounding structure had two jobs: hold the cutting machinery in position and accommodate the particular amino acid sequence the enzyme was meant to cut.

1. RFdiffusion2 designed the protein backbone, constrained to bring the catalytic amino acids together and place the intended target sequence nearby.
2. ProteinMPNN filled that backbone with amino acids chosen to support the planned shape, the catalytic site and attachment to the target.

To check whether the designs worked, the researchers made long proteins containing the candidate enzyme, its intended cutting sequence and a fluorescent protein. Releasing the fluorescent section provided a way to detect a successful cut.

One water molecule changed the results

The cysteine effort produced 13 active designs from 69 tested amino acid sequences, a success rate of 19%. All were less active than the yeast enzyme used as the starting point. The zinc experiments revealed a more specific design problem: water could occupy a position that blocked the reaction site.

14 of 89Without an explicit water-placement constraint

Fourteen of 89 designs showed weak catalytic activity when the design process did not specify the catalytic water molecule.

35 of 38With water placed as a design constraint

Thirty-five of 38 designs showed some catalytic activity when a water molecule’s position was included as a constraint.

Even with that improvement, the zinc designs generally remained considerably less active than their natural starting enzyme. One design had better activity but lacked specificity about where it cut. The experiments therefore exposed a distinction between making a cutter work and making it cut only where intended.

The researchers then tested amino acid substitutions near the active site, avoiding changes predicted to disrupt the structure. Some variants worked substantially better. A revised zinc protease was 30 times more active than its original design, while an improved cysteine protease retained high specificity for its intended sequence.

A cellular cut is not a treatment

For the disease-linked demonstration, the researchers designed zinc proteases against proteins associated with Alzheimer’s disease and frontotemporal dementia. Simply adapting an existing design to a new target sequence worked poorly. Starting the design process again produced an enzyme that cut its intended target in cells.

That result remains an early research demonstration. As Timmer explains in his discussion of therapeutic hurdles, proteins can be difficult to deliver and may trigger immune responses. Reaching a target inside a cell would likely require delivering the gene encoding the protease so the cell could make it.

The design tools also lacked a way to specify a desired level of catalytic activity directly. They could arrange the components and accommodate a target, yet miss subtle features needed for efficient cutting. The gains from subsequent mutations show why matching a planned structure was only part of the engineering task.