The Stability Floor: Broad Institute Shows AI Redesign Unlocks Better Protein Evolution Outcomes

Laboratory evolution works — until the starting enzyme is too fragile to accept the mutations that unlock new function. That stability-activity trade-off has constrained directed evolution for decades: most mutations destabilize proteins, so evolving enzymes toward new substrates often yields variants that fold poorly or never reach the potency of their natural starting points.

A Nature paper published July 24, 2026, from researchers at the Broad Institute of MIT and Harvard offers a practical workaround: redesign the starting sequence with AI before evolution begins.

ProteinMPNN first, PACE second

Led by senior author David Liu, the team used ProteinMPNN — a neural network redesign method from David Baker's laboratory — to stabilize botulinum neurotoxin (BoNT) proteases while preserving catalytic activity. They then ran side-by-side phage-assisted continuous evolution (PACE) campaigns from AI-redesigned starting points versus wild-type enzymes across multiple substrates.

Redesigned starting points consistently outperformed wild-type baselines. In matched BoNT/E campaigns, variants evolved from AI-redesigned backgrounds adapted faster and reached higher activity levels. Critically, mutations that succeeded in redesigned backgrounds often failed when grafted back into wild-type scaffolds — evidence that redesign expanded accessible sequence space rather than merely accelerating stochastic luck.

Ataxin-2: more than 79-fold specificity gain

The therapeutic case study targets ataxin-2, a protein implicated in neurodegeneration. The team evolved BoNT/E protease to cleave ataxin-2 selectively. Proteases evolved from the redesigned starting point achieved more than 79-fold greater selected specificity for ataxin-2 than the best wild-type-evolved enzyme, with higher catalytic efficiency and stability while minimizing cleavage of native substrates.

Rescuing already-evolved enzymes

The workflow also works in reverse. The researchers grafted mutations from a previously PACE-evolved PTEN-cleaving protease — which had lost stability during evolution — onto ProteinMPNN-redesigned backgrounds. The hybrid variants expressed at higher levels in bacteria and mammalian cells and produced more target cleavage product, demonstrating redesign can salvage evolved specificities degraded by stability loss.

Why this matters beyond BoNT

BoNT proteases are a demanding testbed: they are natively part of a delivery holotoxin, with complex specificity constraints and industrial relevance in therapeutics. The paper's broader claim is procedural — AI redesign raises the fitness floor of starting points so continuous evolution spends less time fighting folding failures and more time exploring functional mutations.

That does not eliminate experimental validation. Liu's team still ran dozens of PACE lagoons, kinetic assays, and mammalian cell tests. But it reframes compute as upstream capital in evolution campaigns, not a replacement for them.

For protein engineering teams, the result is a reproducible recipe: stabilize with ProteinMPNN (or related redesign tools), evolve with PACE or analogous selection systems, and expect both faster adaptation and access to sequence neighborhoods wild-type starting points cannot reach.

Sources

• Nature — AI-redesigned starting points and outcomes enhance protein evolution (July 24, 2026)
• Phys.org — Combining AI protein design with laboratory evolution improves engineered enzymes (July 24, 2026)
• News-Medical — AI redesign helps enzymes evolve beyond natural limits (July 24, 2026)