Proteases are enzymes that cut proteins. In two 2026 Nature studies, AI-designed zinc and cysteine proteases showed catalytic activity, and one zinc-protease design reportedly cut its intended target in cells, according to John Timmer’s October 14, 2026, account for Ars Technica. The findings demonstrate activity in experiments, not a treatment; delivery and possible immune responses remain obstacles to therapeutic use.
How AI tools built the designs
The designs started with known enzymes: a zinc protease from a crayfish digestive system and a cysteine protease from yeast. Researchers retained catalytic amino acids—the parts involved in speeding up a chemical reaction—and redesigned the surrounding protein structures.
As Timmer describes it, RFdiffusion2 generated protein backbones under constraints that brought catalytic amino acids together and placed a target sequence near the catalytic center. ProteinMPNN then selected amino-acid sequences suited to the planned backbone, catalytic site, and target binding. AlphaFold3 predicted structures for many designs; researchers reportedly found no close matches in the databases they examined.
A protease cuts peptide bonds, the links that join amino acids into a protein. The design challenge was therefore not just to create a plausible protein shape, but to arrange the reaction site and target so the enzyme could cut the intended sequence.
What the activity screens found
Timmer reports that some candidates in both design efforts showed activity. The screening results were:
| Design set and condition | Reported result |
| Cysteine protease: 12 backbones and 69 amino-acid sets | 13 designs showed activity, reported as 19%; activity was below that of the yeast model enzyme. |
| Zinc protease without a positioned-water constraint | 14 of 89 designs showed weak activity. |
| Zinc protease with a positioned-water constraint | 35 of 38 designs showed some activity; activity generally remained below that of the starting enzyme. |
The two zinc results come from separate design sets with different activity descriptions, not a standardized head-to-head test. The water constraint addressed a specific design problem: modeling of inactive designs suggested that water could block the reaction site. The higher activity count in the separate set is consistent with why the researchers tested that constraint, but the two screens do not establish a controlled effect size.
What the cell experiment means
After attempts to retarget an existing design did not work well, researchers restarted the zinc-protease design process. The resulting enzyme reportedly cut its intended neurological-disease-associated target in cells. The target was associated with conditions including Alzheimer’s disease and frontotemporal dementia.
That is a cell-level result, not a treatment outcome. A therapy would need to reach the right tissue or cells, and the body could mount an immune response. For targets inside cells, delivering a gene that enables cells to make the protease is another hurdle, Timmer reports.
What selected variants changed
Later versions improved on particular designs, not on every candidate. The best reported mutated cysteine-protease variant accelerated cleavage of its intended target by a factor of 10⁹ while remaining highly specific to that sequence. Zinc protease v2 was reported as 30 times more active than the original design and as accelerating peptide-bond cleavage by a factor of 10¹⁰.
Those measurements describe selected variants and their stated comparisons. The reported cell experiment and enzyme-activity results remain research findings; the delivery and immune-response challenges still stand between them and a therapeutic use.
