How AI Gave Protein Engineers a Better Place to Start

Directed evolution has always been a journey with a fixed first step. Researchers pick a natural protein, mutate it millions of times in the lab, and screen for variants that do something useful. The method has produced key reagents in gene editing, antibody drugs, and industrial biocatalysis. The trouble is the starting point. Natural enzymes were not built for the jobs scientists now want them to do, and many of them collapse under the heat, acidity, or solvent conditions that a real therapeutic or manufacturing process demands.

A Nature paper published this week from David Liu's lab at the Broad Institute of Harvard and MIT reframes the bottleneck. The team used a popular AI model to redesign the botulinum neurotoxin, the protein enzyme that gives Botox its activity, as a more stable starting point for laboratory evolution, then ran that redesigned enzyme through standard directed evolution. The evolved variants were substantially more stable and more specific at cleaving a protein linked to neurodegeneration than variants evolved from the natural botulinum neurotoxin. The article frames that substrate as ALS-linked.

The AI's job is to predict mutations that lock the folded structure in place, producing a sturdier first draft that evolution can then push further. After that, the redesigned starting point is handed to the same mutagenesis-and-screening pipeline the field has used for decades. "This insight could change the way researchers conduct protein evolution," Liu said, in a Broad Institute announcement accompanying the paper.

The choice of substrate matters. Botulinum neurotoxin is one of the most dangerous proteins in biology and one of the most useful in medicine, because it can be aimed at a single molecular target. That combination of precision and potency is exactly what a protein engineer wants when trying to silence a misfolded protein implicated in a neurodegenerative disease. The Broad team used it as a stress test: if AI can stabilize a notoriously fragile toxin and the resulting enzyme still cuts a human disease-linked substrate specifically, the same recipe should transfer to less exotic targets.

Liu's lab invented base editing, a CRISPR offshoot that swaps single DNA letters, and has spent years building enzymes that cut or edit proteins with high precision. The new work extends that agenda to the starting point of the evolutionary pipeline. A better starting point expands the universe of reachable proteins, including targets for which no natural enzyme is currently good enough to evolve. In practical terms, the bottleneck is no longer "can nature do this?"; it is "can the AI-redesigned first draft do this well enough for evolution to take it the rest of the way?"

Most aggregator coverage has led with the cosmetic-Botox framing, exemplified by Singularity Hub's piece on why scientists redesigned the Botox enzyme with AI. Phys.org and News-Medical have run the same framing. The Nature paper and the Broad press release are careful on this point. The evolved enzymes cut the ALS-linked substrate in laboratory assays; they have not been shown to work in cells, let alone in patients, and the strategy is being pitched as a way to expand what directed evolution can reach, not as a treatment in itself.

Worth watching next: whether the same AI-stabilization step helps enzymes evolve against other hard targets, whether the redesign model generalizes beyond the protein families it was trained on, and whether the in vitro specificity holds up in animal models. The team is also working on the inverse problem, using AI to redesign the substrate end of a protein pair so an existing natural enzyme can cut something it was never meant to touch.

The model did not design a therapeutic in this paper. It redesigned the first move in a decades-old experiment. Directed evolution does the rest.