AI-Designed Proteases Snip Disease Proteins at Chosen Sites
Scientists at the University of Washington's Institute for Protein Design (IPD) have used artificial intelligence to build proteases from scratch — enzymes that behave like molecular scissors, cutting other proteins at specific points. Two papers describing the work were published in Nature on October 14. David Baker, who directs the institute and shared the 2024 Nobel Prize in Chemistry for computational protein design, is senior author on both. Chen and Kejia Wu are first authors of one paper, and Hojae Choi, an IPD graduate student, leads the other.
Nature spent eons evolving proteases that can cleave other proteins. The researchers short-circuited that process. They kept the catalytic core of two existing protease types and used AI design tools to replace the rest of the protein, producing enzymes that can be tuned to cut a chosen target in half. The designed folds look very different from anything in nature — the cysteine versions score below 0.50 on a standard structural-similarity measure — yet six crystal structures matched the design models closely.
The two studies use different cutting tools. Choi's group built proteases that use an activated amino acid, cysteine, as the blade. Chen's group instead used a zinc ion, which activates a water molecule to attack the peptide bond.
Peptide bonds link amino acids into proteins and can survive for centuries in neutral water, because breaking them takes a great deal of energy. Earlier computational design mostly tackled chemically simpler reactions or artificially activated model systems. The new work takes on one of the most energetically demanding transformations in biology.
Of 69 cysteine-based designs tested, 13 cut their intended targets — a 19% success rate. The most active design accelerated peptide-bond cleavage by up to about 30 million times, compared with the same reaction without an enzyme.
The metalloprotease effort tested 135 designs. Of those, 36% — just over a third — cleaved their cognate substrates at the targeted bond. The best initial design sped up peptide-bond hydrolysis by more than 100 million times; four amino-acid substitutions that improved the active site's preorganization pushed that past 10 billion times. These enzymes selectively cut human disease-associated proteins — TDP-43, amyloid-beta and serum amyloid A — with rate enhancements reaching up to a billion times.
Coverage of the work points to neurodegenerative disease. The IPD enzymes target proteins tied to Alzheimer's disease and amyotrophic lateral sclerosis (ALS), and they cut over and over: once the enzyme destroys one target molecule, it can move on to the next. Baker's team says it is now pursuing targets for malaria, chronic diseases and cancer, and that similar proteins could disarm viruses or bacterial toxins.
In the lab, the designs catalyze protein breakdown in mammalian cells. That is a long way from a treatment for people, and the researchers do not claim one. The bigger shift may be conceptual: scientists can now order up custom enzymes instead of relying on naturally evolved proteases. Baker frames the ambition through AI BioDesign, a project run with the Allen Institute and the Fred Hutchinson Cancer Center that aims to design molecules nature never made — his stated goal being a cure for a new disease in weeks rather than decades.