AI-Redesigned Enzymes Evolve Beyond Natural Limits, Enabling Therapeutic Protein Engineering

A team of researchers has demonstrated that artificial intelligence can redesign enzymes in ways that unlock evolutionary pathways nature never developed. The work, published in Nature, centers on botulinum neurotoxin proteases—enzymes that cleave specific proteins—and shows that when you stabilize these molecular machines before subjecting them to directed evolution, they can access mutations that would have been toxic or non-functional in their original form.

The problem the researchers were solving is fundamental to protein engineering. Wild-type enzymes, the ones that exist in nature, often operate under tight constraints. They have just enough stability to do their job, which means that when you try to evolve them toward new functions—making them cleave different proteins, for instance—the mutations that confer those new abilities tend to destabilize the enzyme's structure. The enzyme gains function but loses integrity. It's a trade-off that has long limited what directed evolution can achieve.

The team used two computational tools to sidestep this constraint. ProteinMPNN, a deep-learning model, redesigned the amino-acid sequences of botulinum proteases while preserving their three-dimensional structure. A second tool called PROSS added further stabilization. The result was a set of redesigned enzymes that were thermally more robust—some melting at temperatures up to 59.5 degrees Celsius compared to the wild-type baseline—and paradoxically, many were already more catalytically efficient than their natural counterparts, with some showing 1.7 to 2.8 times greater catalytic efficiency in initial tests.

Then came the evolution phase. The researchers subjected both the redesigned enzymes and wild-type versions to phage-assisted continuous evolution, a high-throughput process that screens billions of variants in parallel. They ran 44 separate evolution campaigns on an automated platform, challenging the enzymes to cleave increasingly difficult substrate targets. The results were stark. When the most challenging substrate was introduced, wild-type evolution campaigns failed in half of the experimental runs. The redesigned enzymes succeeded in all of them. More tellingly, the redesigned variants tolerated mutations—like one called K225E—that would have completely disabled the wild-type enzyme but that conferred high catalytic function in the stabilized background.

The researchers then applied this approach to a therapeutic target: human ataxin-2, a protein implicated in neurodegeneration and ALS. They evolved one of their redesigned proteases to cleave ataxin-2 instead of the enzyme's native substrate, SNAP25. The evolved variant achieved more than 79-fold greater specificity for ataxin-2 compared to the best wild-type-derived enzyme, and remarkably, it showed no detectable cleavage of the original SNAP25 substrate even at high concentrations. The protein had been reprogrammed. It was 16 percent divergent from the natural framework, yet it worked.

What makes this work significant is not just the immediate results but what it reveals about the relationship between stability and evolvability. By giving enzymes more structural slack—more thermodynamic headroom—the researchers expanded the landscape of mutations that could be tolerated. Mutations that would have been evolutionary dead ends in a marginally stable enzyme became viable stepping stones in a stabilized one. A kinetically impaired redesign, one that started out 20 times slower than wild-type, eventually evolved to higher activity than the natural enzyme. The starting point mattered more than the starting speed.

The implications extend beyond botulinum proteases. The workflow demonstrates a general principle: computational stabilization followed by directed evolution can accelerate the engineering of enzymes for therapeutic use. If you can design proteins that are stable enough to explore a wider mutational space, you can evolve them toward disease targets that might otherwise remain out of reach. The approach could reshape how researchers approach enzyme engineering for biomedical applications, environmental remediation, and industrial biocatalysis. The question now is how broadly this principle applies and how quickly it can move from the laboratory into clinical development.