AI Redesign Helps Enzymes Evolve Beyond Natural Limits

Researchers used AI-based protein sequence redesign to create stabilized botulinum neurotoxin (BoNT) protease variants that evolve more efficiently than wild-type enzymes. According to News-medical, this workflow allowed a redesigned protease to achieve over 79-fold greater specificity for human ataxin-2, a protein linked to neurodegenerative diseases like ALS.

Engineering enzymes often hits a wall known as the stability-activity trade-off. In conventional directed evolution, mutations that grant a protein new functions or a better ability to cleave specific substrates often make that protein thermodynamically unstable. The enzyme essentially falls apart before it can reach peak performance.

To bypass this, a study published in the journal Nature integrated artificial intelligence with high-throughput evolution platforms. By stabilizing the “starting point” of the enzyme before the evolution process begins, the team opened mutational routes that were previously inaccessible to natural, wild-type proteins, according to News-medical.

ProteinMPNN and PROSS Sequence Redesign

The researchers utilized two specific computational tools to prepare their enzyme candidates: a deep-learning sequence-design model called ProteinMPNN and a protein-stabilization method named PROSS. These tools were used to redesign the catalytic domains of BoNT/E, BoNT/F, and BoNT/X.

These redesigns weren’t random. The team applied multiple-sequence alignment (MSA) conservation thresholds between 30–60% and incorporated structural distance constraints of 10–18 Å from the catalytic zinc ions and the substrate, News-medical reports.

The results of the initial redesign phase for BoNT/E were promising. Out of 74 ProteinMPNN designs, 78% maintained their catalytic activity. Some of the top variants, labeled D1-D3, showed catalytic efficiencies 1.7 to 2.8 times higher than the wild-type enzyme.

Boosting Catalytic Efficiency and Thermal Stability

The shift from wild-type to AI-redesigned sequences produced measurable gains in how these enzymes operate. One specific variant, D2, reached a catalytic efficiency (kcat/KM) of up to 310 mM−1s−1, a significant jump from the 110 mM−1s−1 recorded for wild-type BoNT/E, according to News-medical.

Stability also improved. The redesigned enzymes demonstrated melting temperatures reaching up to 59.5°C. This increased thermal stability is the key to the study’s success; by ensuring the enzyme doesn’t unfold under pressure, the researchers could push the protein toward more extreme and specific functions without the structure collapsing.

The eVOLVER Platform and PACE Evolution

Once the stabilized starting points were created, the researchers subjected them to phage-assisted continuous evolution (PACE) using an automated eVOLVER platform. They ran 44 parallel evolution campaigns, challenging both wild-type and redesigned BoNT/E proteases against a series of altered SNAP25 substrates (specifically substrates 412, 413, and 415) that increased in difficulty.

The AI-redesigned variants didn’t just perform better; they evolved faster. News-medical reports that these redesigned starting points accessed a highly active mutational space that was completely non-functional when using the wild-type BoNT/E background.

Targeting Human Ataxin-2 in Neurodegeneration

The ultimate test of this workflow was the attempt to reprogram BoNT/E to target human ataxin-2 (residues 1181–1201), a protein implicated in ALS and other neurodegenerative conditions. The contrast between the AI-assisted approach and the traditional approach was stark.

The AI-redesigned protease variant achieved more than 79-fold greater specificity for the ataxin-2 substrate than the best enzyme evolved from the wild-type. This variant showed a 16% sequence divergence from the natural protein framework, proving that AI can push an enzyme far beyond its natural evolutionary limits to create a highly specific tool.

HEK293T Cell Expression and PTEN Cleavage

The utility of these redesigns extended to cellular expression. By combining ProteinMPNN redesigns with mutations from a previously PACE-evolved PTEN-cleaving BoNT/E protease, the researchers saw a massive increase in HEK293T cell expression. Specifically, the D2 and D3 variants increased expression by over 24-fold.

This boost in expression translated to higher activity: PTEN cleavage products increased by 4.5-fold for D2 and 3.9-fold for D3, according to News-medical.

This development indicates that the primary bottleneck in therapeutic enzyme design is often not the final goal, but the fragility of the starting material. By using AI to build a “sturdier” protein from the outset, researchers can navigate complex fitness landscapes that would otherwise be lethal to the enzyme’s structural integrity.