AI news · October 15, 2026
AI-designed proteases cut targeted proteins in lab studies
The model stack is moving from structure prediction to designed function, but delivery, immune response and activity outside the lab remain the hard gates.
- zinc designs worked
- 35/38
- best acceleration claim
- 10^10
- cysteine design success
- 19%
Ars Technica reported on October 14 on work from the University of Washington and the David Baker lab that uses AI to design proteases, enzymes that cut other proteins. RFdiffusion2 generates a backbone, ProteinMPNN assigns amino acids and AlphaFold3 predicts structures. The reported experiments produced working zinc-dependent designs in 35 of 38 candidates, while the best variant accelerated cleavage by a claimed ten billion times. Early cell tests targeted proteins linked to Alzheimer’s and frontotemporal dementia.
This is a meaningful change from predicting a natural molecule to designing one for a chosen function. It is not a treatment. The report notes delivery, immune response and the gap between predicted structure and useful activity as major hurdles. For builders, the transferable pattern is a closed loop: generate, predict, test and discard quickly. Keep biological claims tied to the experiment, not to the model name.
What you can do with it
For a research workflow, separate design score from wet-lab success. Store the generated sequence, predicted structure, assay result and failure reason for every candidate. Promote a design only when it passes the next physical test, not when the prediction score is high.
Our take
The result shows how model composition can create a new search space for biology. The commercial path is still long because function, delivery and safety must all hold. The best near-term opportunity is faster experimental triage, not a shortcut around validation.
Source: Ars Technica — Made With Models writes the brief; the reporting is theirs.