AI-Designed Botox Enzyme Cuts ALS Protein 80 Times Faster Than Natural Counterparts

Artificial intelligence is reshaping the way scientists tailor enzymes, the biological catalysts that drive virtually every cellular reaction. In a recent study, a team led by David Liu at the Broad Institute used an AI‑based design tool to remodel the protease component of Botox, creating variants that more efficiently cleave a protein fragment implicated in amyotrophic lateral sclerosis (ALS).

Engineering new enzymes has traditionally relied on “directed evolution,” a labor‑intensive process that mutates a natural protein over many generations while selecting for desired traits. The method can span months and still fail if the starting molecule is too fragile to tolerate further changes.

“The quality of the initial scaffold dictates the outcome of laboratory evolution,” Liu explained in a press release. “Investing time in a robust starting point can dramatically reduce the resources needed for downstream optimization.”

Natural enzymes often collapse under the strain of repeated mutations, limiting the breadth of reactions they can support. Enhancing their intrinsic stability before evolution begins could unlock a wider array of biochemical functions.

AI‑Generated Blueprints Bypass Traditional Limits

Liu’s group turned to ProteinMPNN, a deep‑learning model developed by David Baker’s lab, which can propose novel amino‑acid sequences that preserve a protein’s overall fold. By feeding the botulinum neurotoxin protease into the algorithm, the researchers obtained 58 candidate designs predicted to be more stable than the wild‑type enzyme.

Three of the top candidates were expressed in Escherichia coli and displayed high solubility, a key indicator that they remained properly folded. In several cases, activity measurements even exceeded those of the native protease.

These AI‑engineered enzymes were then subjected to Phage Assisted Continuous Evolution (PACE), the high‑throughput platform Liu introduced in 2011 that accelerates evolutionary cycles by linking protein performance to bacteriophage replication. The goal was to improve cleavage of a mutant segment of the protein TDP‑43, whose pathological expansion drives neurodegeneration in ALS.

When benchmarked against PACE‑evolved variants derived from the natural toxin, the AI‑originated enzymes achieved roughly an 80‑fold increase in catalytic efficiency and demonstrated more than 56‑fold greater selectivity for the target site.

“Starting from a protein that already possesses excess stability gives the evolutionary process room to explore more radical mutations without sacrificing function,” said study co‑author Nicholas Krasnow.

Implications and Next Steps

Although the experiments were conducted in bacterial and immortalized human cell systems, the findings suggest that coupling AI‑driven design with continuous evolution could streamline the creation of bespoke enzymes for therapeutic and industrial applications. Liu’s lab is already extending the workflow to refine prime editors and other genome‑editing tools.