First Stabilize the Enzyme Scaffold: AI Design Opens a New Path for Laboratory Evolution

Engineering therapeutic enzymes often involves a difficult trade-off: mutations introduced to hit a new target may instead cause the protein to lose stability or activity. A study published in *Nature* proposes a two-stage strategy, first using artificial intelligence to redesign a protein “scaffold” that can better tolerate changes, then using continuous laboratory evolution to identify sequences with new functions.

The research team used ProteinMPNN to redesign a botulinum neurotoxin protease, followed by phage-assisted continuous evolution. This method repeatedly replicates and selects phages carrying different protease sequences, gradually retaining variants that can more effectively recognize and cleave a designated substrate. The study compared four evolution experiments starting from either AI-redesigned proteins or wild-type proteins.

The results showed that the two evolutionary routes did not merely reach the same endpoint at different speeds. Starting from the AI-redesigned versions, the researchers obtained a set of well-functioning sequences; some key mutations, when transferred directly into the natural protease background, failed to function properly. The team therefore hypothesized that the stability introduced by redesign acts as additional structural leeway, allowing the protein to accommodate mutations that would otherwise be too costly but help alter its function.

In a demonstration more closely related to disease research, the team evolved a stabilized protease to cleave an aggregation-associated region of ataxin-2. Ataxin-2 is associated with mechanisms of neurodegenerative disease; versions evolved from the AI-redesigned protease not only showed greater catalytic efficiency and stability, but also had more than 79-fold higher “selected specificity” for ataxin-2 than the best variant obtained by starting from the natural enzyme.

This difference supports a practical concept: AI protein design may not need to predict the final drug in a single step. Instead, it can first create a starting point better suited to evolution, then allow high-throughput selection to explore a complex functional landscape. In this way, evolution is no longer constrained by the original stability limits of natural proteins and may also identify combinations of mutations that cannot function in a wild-type background.

However, this remains a proof of concept in protein engineering and cannot be regarded as evidence that an ataxin-2 therapy has demonstrated clinical efficacy. The more than 79-fold figure reflects specificity under particular experimental selection conditions and cannot yet be directly translated into efficacy or a safety range in humans. Future studies will still need to evaluate delivery in cells and animals, unintended cleavage, immune responses, and long-term toxicity. PubMed data also disclose patent applications associated with the Broad Institute, indicating that this method has translational and commercialization potential and making independent validation of subsequent findings all the more important.