AI Designs Proteases That Recognize and Cut Specific Protein Targets

Researchers at the University of Washington, led by David Baker, have used artificial intelligence to design working proteases, molecular scissors capable of recognizing and cutting selected protein targets. The work, reported in Nature, marks a shift in AI-driven protein design from generating new shapes and structures to creating functional enzymes that perform targeted biological tasks. The team developed a pipeline that combined several advanced AI models to design both cysteine and zinc proteases from scratch. Using RFD2-MI, a version of RFdiffusion2, the researchers generated three-dimensional protein shapes and backbone cores. ProteinMPNN and EnhancedMPNN then selected amino acid sequences intended to make the enzymes fold correctly and bind their chosen targets. AlphaFold3 was used to assess whether the final structures matched the design goals, and molecular dynamics simulations helped evaluate the intended cutting mechanism. The designs were tested in bacteria by producing linked chains that combined candidate enzymes, target sequences and a glowing marker protein. A design was considered successful if the enzyme cleaved the target and released the marker. Of 69 cysteine protease candidates, 13 showed activity. The zinc protease designs performed more efficiently, with 36% of 135 candidates cutting their matching peptide targets. Further AI refinement increased the cutting speed of the leading zinc protease by 30 times and produced a cysteine protease that cut precisely at the intended site. The study shows that de novo enzyme design has moved beyond simple model reactions and can generate proteases with practical target recognition and catalytic activity. Some of the engineered enzymes also suggested therapeutic potential. One protease cut full-length TDP-43, a protein linked to ALS and some types of dementia, while other designs targeted fragments of amyloid-beta associated with Alzheimer’s disease and serum amyloid A linked to systemic amyloidosis. Although the enzymes have not yet been tested in human clinical trials, the results point to a broader impact for AI-designed proteins. The ability to create selective proteases for chosen protein regions could support new approaches in medicine, biotechnology and industrial chemistry.