2026-07-27
Using ProteinMPNN-redesigned botulinum proteases as directed-evolution starting points consistently yields higher-activity enzymes across four campaigns, with 79-fold higher selectivity for ataxin-2 than wild-type starts.
Engineered or lab-evolved enzymes often have suboptimal stability, activity, or specificity. The usual route is to start from the wild-type (WT) protein and stack mutations through directed evolution, but the natural starting point may sit far from the functional peak you want, with a valley of unfavorable mutations in between. This Nature paper tries a counterintuitive move: first redesign the sequence with AI and use the redesigned variant as the evolution starting point, instead of the natural WT.
The model is ProteinMPNN, the sequence-design network from the University of Washington. The logic: an MPNN-redesigned enzyme is more structurally regular and stable, and likely more mutationally robust, meaning nearby mutations survive more easily. If so, it is a better launchpad for evolution.
The team first used ProteinMPNN to redesign three botulinum neurotoxin (BoNT) proteases, obtaining variants with improved stability and full catalytic efficiency. Then they ran a head-to-head: under identical selection pressure, they started evolution from the AI-redesigned protease or from the corresponding WT, using PACE (phage-assisted continuous evolution), a high-throughput method that lets a protein accumulate many mutations over days. Four evolution campaigns compared which starting point yields higher-activity enzymes.
Across four campaigns, evolution from the AI-redesigned starting point consistently produced proteases with higher activity than evolution from WT. A key piece of evidence: mutations that evolved on the redesigned background stop working when moved back into the WT background. This shows the redesigned starting point reached a highly functional region of sequence space that WT cannot reach.
The most concrete result: they evolved both natural BoNT/E and AI-redesigned BoNT/E to selectively cleave ataxin-2, a therapeutically relevant target tied to spinocerebellar ataxia. The variant from the redesigned starting point reached higher catalytic efficiency and stability while minimizing cleavage of the native substrate, achieving more than 79-fold greater selectivity for ataxin-2 than the best variant evolved from WT.
This gives AI protein design a clear new role: a stronger starting point for experimental evolution, beyond just generating end products. For protein engineering and synthetic biology, enzymes that will not budge from their natural form can first be moved by AI to a more evolvable position, then finished by directed evolution. The workflow is general, demonstrated across three proteases.
Access to this Nature article was at the abstract level; the specific body numbers (per-variant Tm thermal stability, kcat/KM catalytic constants, mutation counts) were not retrievable from the accessible text, so the account above is bounded by what the abstract discloses. At that level, mutational robustness is argued indirectly (mutations fail when moved to WT) rather than measured as a direct metric. The demonstration is concentrated on three BoNT proteases; whether the workflow transfers to other enzyme families needs more validation.