A team led by David Liu at the Broad Institute of MIT and Harvard has used AI to create more stable versions of the botulinum neurotoxin enzyme-the active component in Botox-and then evolved them to precisely cut a protein associated with ALS. The approach produced enzymes that were 80 times more efficient and 56 times more selective than those evolved from natural enzymes, offering a new way to engineer proteins for research and medicine.
The Limits of Directed Evolution
Enzymes are proteins that speed up chemical reactions, and researchers often modify them for gene editing, synthetic biology, and therapeutics. The standard method, directed evolution, starts with a natural enzyme and introduces mutations over many generations to select for desired traits. But natural enzymes can be fragile: even small changes can cause them to lose stability and stop working. This makes the starting enzyme a critical factor in the success of the process.
"Laboratory evolution requires the commitment of time and resources. So what you start with is incredibly important as a major determinant of what you end up with," said Liu in a press release.
Liu's lab previously developed a system called PACE that automates directed evolution, running dozens of rounds per day. Yet they found that almost all successes began with natural proteins, and these often became unstable as they evolved new functions. The resulting enzymes could clump together, lose precision, or still act on their original targets. Workarounds, such as adding chaperone proteins or pre-stabilizing enzymes, added complexity and time.
AI Redesigns the Enzyme
The team turned to an AI model called ProteinMPNN, developed by David Baker's group at the University of Washington. ProteinMPNN can generate new protein sequences that maintain a desired structure while altering the amino acid building blocks-a process that takes seconds. The researchers used it to redesign the botulinum neurotoxin protease, aiming to increase its stability without changing its function. This approach is part of a broader trend in AI for science and research.
ProteinMPNN produced 58 designs predicted to be more stable. The top three candidates, when produced in E. coli, were highly soluble and some even showed higher activity than the natural enzyme. These AI-redesigned enzymes served as starting points for directed evolution using PACE.
Superior Performance Against an ALS Target
The team evolved the enzymes to cut a mutated region of a protein linked to neuron health. In diseases like ALS, a repetitive stretch in this protein expands, causing it to aggregate and damage neurons. Naturally occurring enzymes have struggled to cut the mutant version before it forms toxic clumps.
When compared to enzymes evolved from the natural botulinum neurotoxin, those derived from the AI-redesigned versions were nearly 80 times more efficient at cutting the target and over 56 times more selective for the intended region. This held across three types of the neurotoxin and multiple substrates. Mathematical mapping of the evolutionary paths showed that the redesigned enzymes tolerated more mutations while gaining new functions.
"If you start with a more stable protein, it has more stability to spare, so it can afford larger changes in pursuit of new functions," said study author Nicholas Krasnow.
Why this matters for science and research professionals
The study demonstrates that AI-generated protein designs can serve as superior starting points for directed evolution, potentially expanding the range of targets that enzyme engineering can address. The experiments used immortalized human cells, so further testing in more complex systems is needed. For researchers working on protein-based therapeutics, diagnostics, or synthetic biology, this approach could reduce the time and labor required to develop enzymes with novel functions. The team is already applying the strategy to improve prime editors and other molecular tools, signaling that the method may soon influence a variety of biotechnology workflows.
Your membership also unlocks:
