Building a Stable Foundation for Proteins First: AI Takes Laboratory Evolution Further

Protein engineering often involves a difficult trade-off: to teach an enzyme to recognize a new target, researchers must accumulate mutations, but these can also make it unstable, difficult to fold, or even unable to be sufficiently expressed in cells. The Broad Institute team’s solution is to first use artificial intelligence to give the protein a stable structural foundation, then let laboratory evolution explore new functions.

This proof-of-concept study, published in *Nature*, used botulinum neurotoxin proteases as a model. The team used ProteinMPNN to redesign their amino acid sequences while preserving the regions required for catalysis and substrate binding, and then used phage-assisted continuous evolution (PACE) to select variants capable of cleaving specified proteins. PACE links enzyme activity to phage replication, allowing sequences that better meet the target criteria to accumulate rapidly over multiple rounds of mutation and selection.

Not all initial designs succeeded. Of 74 BoNT/E protease designs, 58 remained functional; among them, 22 both retained activity and produced higher yields of soluble protein than the wild type. The researchers then compared the performance of AI-redesigned versions and natural enzymes as starting points for evolution. In tests involving multiple botulinum proteases and different substrates, the redesigned starting points consistently produced evolved outcomes with higher activity.

The test with the greatest disease relevance focused on ataxin-2. This protein is associated with the pathological processes of some neurodegenerative diseases, and the team attempted to make the protease selectively cleave a specific region of it. The best variant evolved from an AI-redesigned starting point showed more than a 79-fold increase in “post-selection specificity” for ataxin-2 compared with the best version evolved from natural BoNT/E. It also had better stability, was expressed at higher levels in mammalian cells, and produced more ataxin-2 cleavage products.

Sequence differences suggest that stability may provide a buffer that accommodates mutations. The leading variant accumulated 64 mutations and differed in sequence from the natural enzyme by about 16%; by comparison, the best variant derived from the natural starting point had only 10 mutations and differed by about 2.4%. Some mutations that increased the activity of the redesigned protein caused destabilization when transferred back into the natural protein background, supporting the explanation that “stability headroom must first be increased before a greater functional distance can be traversed.”

However, this result is not yet a therapy for neurodegenerative diseases. The study primarily demonstrates a protein-engineering workflow, and the existing evidence comes from in vitro analyses, bacterial expression systems, and cultured cells. It has not yet answered how the protease could be safely delivered into target tissues, whether it would cleave other human proteins, whether it could improve disease in animals, or how immune and safety risks associated with an enzyme derived from botulinum toxin would be addressed. If this strategy can be reproduced in other protein classes, it may in the future be used to improve therapeutic enzymes and gene-editing tools, but each application will still need to be validated separately.