AI-Redesigned Botox Enzyme Shows Promise for ALS Treatment

Protein engineering has always been a game of patience. Scientists take a natural enzyme—one of the body's molecular workhorses, responsible for everything from cutting DNA to breaking down toxins—and nudge it toward new abilities through a process called directed evolution. It's tedious work. Each generation takes time. Each mutation risks collapse. And after months or years of careful iteration, success is never guaranteed.

The bottleneck has been stubborn. When researchers evolve enzymes in the lab, the proteins often become unstable as they mutate. They lose their shape. They clump together inside cells and stop working. Even when evolution succeeds in creating an enzyme that can recognize a new target, it may still fumble its original job, cutting the wrong protein by accident. Natural enzymes, it turns out, make imperfect starting points for the kind of radical redesign that medicine needs.

David Liu and his team at the Broad Institute of Harvard and MIT have spent years wrestling with this problem. In 2011, they built a system called PACE that could run dozens of evolutionary rounds per day, dramatically accelerating the process. Using it, they engineered more efficient gene editors, precise RNA-targeting enzymes, and therapeutic antibody fragments. But even PACE hit a wall. Nearly every success began with a natural protein, and nearly every evolved descendant grew fragile.

So they turned to artificial intelligence. Over the past decade, AI models trained on protein structures have become powerful enough to design entirely new sequences that preserve a protein's overall shape while changing its molecular building blocks—all in seconds. Liu's team used ProteinMPNN, a model developed by Nobel laureate David Baker at the University of Washington, to redesign the enzyme behind Botox. The botulinum neurotoxin protease is a molecular scissor that paralyzes muscles by cutting specific proteins. ProteinMPNN generated 58 candidate designs predicted to be more stable. When the team produced the top three in bacteria, they were highly soluble and didn't aggregate. Some even outperformed the natural version.

Then came the real test. The team fed these AI-redesigned enzymes into PACE and evolved them to cut a mutated protein linked to ALS. In that disease, a repetitive stretch of amino acids expands, causing the protein to clump and gradually destroy neurons. Natural enzymes have struggled to cut this target before the damage sets in. But enzymes descended from the AI-redesigned versions were nearly 80 times more efficient at slicing the mutant protein, and over 56 times more selective for the intended region. Across three different types of botulinum neurotoxin and multiple substrates, the AI-designed starting points consistently produced more stable, more effective enzymes.

The reason is structural flexibility. When you begin with a more stable protein, it has room to spare—it can tolerate larger mutations while still holding together and gaining new functions. By mapping the evolutionary paths mathematically, the team found that the redesigned enzymes could afford changes that would have broken their natural counterparts. This extra tolerance opens a door that has been closed for decades: the possibility of engineering enzymes for protein targets that have no suitable natural enzyme to start with.

The work so far has been done in immortalized human cells, so whether these proteins will perform as well in living tissue remains an open question. But the principle is clear. Coupling AI design with laboratory evolution doesn't just speed up the process—it fundamentally changes what's possible. Liu's team is already applying the strategy to other molecular tools, including prime editors. The insight, as Liu put it, could reshape how researchers conduct protein evolution. For diseases like ALS, where time is the enemy and current treatments are limited, that shift in approach might matter enormously.