AI Creates Custom Cell Hunters To Destroy Incurable Diseases

Researchers have used generative AI to design entirely new protein-cutting enzymes from scratch, overcoming a longstanding challenge in computational biology. The synthetic enzymes, known as proteases, can recognize and cut specific sequences of amino acids, potentially giving scientists a powerful new way to manipulate biological processes, and eventually, develop treatments for diseases caused by proteins, including Alzheimer's, Parkinson's, and even type-2 diabetes.

The research, published in Nature yesterday and titled "De novo design of cysteine proteases", was led by scientists at the University of Washington who used AI protein-design tools to generate three-dimensional protein structures and amino-acid sequences capable of performing the necessary chemistry. Proteases work by breaking peptide bonds, the chemical links that hold proteins together. Although nature is full of enzymes capable of doing this—and doing it better than these synthetic ones—designing entirely new proteases has proved exceptionally difficult because peptide bonds are remarkably stable and require carefully positioned chemical machinery to break efficiently.

Of the 69 initial designs tested, 13 successfully cleaved their intended targets. The researchers then optimized the most promising candidates, eventually producing a variant called Dokki-v2. When benchmarked against the reported rate of peptide-bond hydrolysis without an enzyme, the optimized design accelerated the reaction by a rate of approximately 1.8 billion times, which sounds impressive, but it's still about 2.5 times slower than natural enzymes. The breakthrough is that it worked at all; the result demonstrates that AI-designed proteins can do substantially more than simply fold into shapes that look plausible on a computer screen.

The researchers also confirmed that the designed enzymes adopted the intended structures. Six X-ray crystal structures closely matched their computational models, with differences of less than 1.2 angstroms—a measure that should be familiar to HotHardware readers given its recent use in chip manufacturing. The new enzymes also had structural folds distinct from those found in natural proteases. Importantly, experiments apparently showed that at least one extensively tested design could distinguish its intended peptide sequence from seven similar alternatives, an essential capability for applications where cutting the wrong protein could be harmful.

Arguably the most significant finding is that members of the Dokki enzyme family were proven to operate inside living human cells. In experiments involving HeLa cells, the enzymes successfully cut an engineered reporter protein, causing a fluorescent signal to move away from the cell nucleus. That's a critical proof of concept, but it's just one step; scientists still have to prove that the enzymes can reach particular tissues, find disease-associated proteins, and safely destroy the unwanted material inside a living human being without causing mass chaos elsewhere in the body.

It was the authors who suggested that computationally designed proteases could eventually be tailored to target disease-related proteins, or to enable even more sophisticated forms of cellular engineering, but this is very much a research win, not a product-stage breakthrough. The most active AI-designed proteases still lag behind naturally occurring proteases in terms of catalytic turnover, after all. This is a demonstration of a promising new protein-engineering technique, not a ready-made medical treatment. Whether it can be developed into an effective therapy will depend on what researchers can successfully target, and whether they can do so safely in actual humans.