Stability First: ProteinMPNN Redesigns Are Rewriting the Starting Line for Enzyme Evolution

The bottleneck was never the selection—it was the starting point

Directed evolution has been one of biotechnology's most reliable engines for turning natural enzymes into tools with new functions. Phage-assisted continuous evolution (PACE), pioneered in David Liu's laboratory at the Broad Institute of MIT and Harvard, can run dozens of mutation-and-selection generations in a single day without manual intervention. The method has reprogrammed DNA-binding proteins, genome editors, and proteases with specificity shifts exceeding a million-fold.

Yet a stubborn pattern keeps appearing in the results: evolved enzymes often arrive with compromised stability, lower catalytic potency on their new substrates, and lingering activity on the targets evolution was meant to leave behind. According to the authors of a study published in Nature on July 24, 2026, the problem is not that laboratory evolution fails—it is that natural proteins start the race with too little folding margin to absorb the destabilizing mutations that new functions demand.

The Broad team, led by graduate student Nicholas A. Krasnow, proposes a practical fix: redesign the enzyme sequence with AI before evolution begins. Using ProteinMPNN—the neural network sequence-design model from David Baker's laboratory—they stabilized botulinum neurotoxin (BoNT) proteases and then ran side-by-side PACE campaigns against wild-type starting points. In every matched comparison, the AI-redesigned enzymes evolved to higher activity, adapted faster, and unlocked mutation paths that simply did not function in natural backgrounds.

Why natural enzymes run out of evolutionary headroom

Most wild-type proteins sit near the edge of their folding stability. Because the majority of random mutations are destabilizing, an enzyme accumulating changes to gain a new function often loses the ability to express or fold cleanly long before it reaches peak activity. Liu's group has documented this trade-off across PACE selections for proteases, base editors, and other reprogrammed catalysts: specificity can shift dramatically, but catalytic efficiency on the new substrate frequently lags what the native enzyme achieved on its original target.

Researchers have tried workarounds—pre-evolving for thermal stability, overexpressing chaperones during selection, or rescuing degraded variants after the fact. Each adds time, labor, or dependency on conditions that may not transfer to therapeutic manufacturing. Krasnow and colleagues asked whether computational redesign could front-load stability instead.

ProteinMPNN and the physics-based PROSS method have both demonstrated that natural protein sequences can be rewritten for greater stability while preserving native function. What remained untested at scale was whether those redesigned sequences would serve as better evolutionary starting points—not just better end products—across diverse substrates and selection pressures.

Building stabilized proteases with ProteinMPNN

The team chose BoNT proteases as their model system. These enzymes—familiar from their role in Botox therapeutics—naturally cleave neuronal SNARE proteins. Liu's lab had previously used PACE to reprogram BoNT proteases toward non-native targets with large specificity gains, making them a rigorous testbed for the redesign hypothesis.

For each BoNT serotype, the researchers constrained residues critical to catalysis, zinc coordination, and substrate binding, then allowed ProteinMPNN to redesign the remaining sequence against the wild-type crystal structure. Candidate designs were filtered with AlphaFold2 structure predictions before experimental screening.

The numbers from the initial BoNT/E campaign are instructive. Among 74 ProteinMPNN designs assayed, 78% retained catalytic activity, and 45% matched or exceeded the wild-type cleavage rate on the native SNAP25 substrate. The top three designs (D1–D3) showed 1.7- to 2.8-fold higher catalytic efficiency than wild-type BoNT/E, with D2 reaching a kcat/Km of 310 mM⁻¹ s⁻¹ compared with 110 mM⁻¹ s⁻¹ for the natural enzyme. They also expressed more solubly in E. coli and melted at higher temperatures.

Similar redesign-and-characterization workflows succeeded for BoNT/F and BoNT/X proteases, demonstrating that the approach generalizes across serotypes with distinct native specificities.

Side-by-side evolution: redesigned starters win every time

The central experiment paired AI-redesigned BoNT/E protease D3 against wild-type BoNT/E across four evolution campaigns on three substrates of increasing difficulty. Using autonomous eVOLVER devices to run 22 independent PACE lagoons in parallel, the team evolved each starting point toward new cleavage specificities.

The pattern held consistently:

• On the easiest substrate, D3-evolved proteases reached 6–8-fold higher circuit activation than wild-type-evolved counterparts in matched replicates.
• On intermediate-difficulty substrates, redesign-evolved variants achieved up to 15-fold activation versus 8.5-fold for wild-type-evolved enzymes—and adapted with fewer mutations because functional single mutants survived selection earlier.
• On the hardest substrate, two of four wild-type replicates failed entirely before phage washout, while all four D3 replicates succeeded.

Critically, mutations that conferred high activity in the redesigned background often failed when grafted into the wild-type enzyme. Genotype-rescue experiments showed that wild-type proteases required compensatory stabilizing mutations to reach the same activity levels that redesigned backgrounds achieved directly. Long-read sequencing of evolved populations confirmed that redesign-specific mutation sets propagated in redesign-initiated lagoons but rarely took hold in wild-type-initiated pools.

As Krasnow put it in a Broad Institute release accompanying the paper: when proteins evolve new functions, they typically sacrifice stability—and that caps how far they can change. Starting with a more stable scaffold gives evolution stability to spend.

Rescuing evolved enzymes—and aiming at ataxin-2

The workflow is not limited to fresh starting points. The team grafted 16 PACE-evolved mutations from a previously reprogrammed PTEN-cleaving BoNT/E variant into ProteinMPNN-redesigned backgrounds. Expression in bacteria improved up to 5.2-fold; in human HEK293T cells, redesigned grafts expressed at more than 24-fold higher levels while preserving reprogrammed specificity and eliminating detectable cleavage of the native SNAP25 substrate.

The therapeutic headline comes from evolving BoNT/E toward ataxin-2, a protein implicated in neurodegenerative disease. Proteases evolved from the AI-redesigned starting point achieved higher catalytic efficiency and stability while minimizing off-target cleavage of SNAP25. The best redesign-derived variant showed more than 79-fold greater selected specificity for ataxin-2 than the top performer evolved from wild-type BoNT/E.

That figure matters because substrate specificity—not raw catalytic rate alone—determines whether a reprogrammed protease can be deployed as a precision therapeutic rather than a blunt instrument.

Neither AI nor evolution alone—and what comes next

The study does not claim ProteinMPNN replaces directed evolution, or that PACE makes sequence design obsolete. The authors argue the combination addresses limitations inherent to each method: de novo-designed enzymes still lag natural catalysts in raw turnover, while evolution from natural starting points often stalls on stability cliffs.

David Liu, senior author and director of the Merkin Institute of Transformative Technologies in Healthcare at the Broad, stated in the institute's announcement that using AI to stabilize natural proteins could change how researchers conduct protein evolution—a shift from treating stability as an afterthought to treating it as the first engineering step.

The team is now applying the strategy broadly, including to reverse transcriptases whose stability bottlenecks prime editing systems. Whether the workflow transfers cleanly to enzyme classes beyond BoNT proteases remains an open empirical question; the authors note that broader applicability requires future testing.

For the biotech industry, the implication is operational. As AI protein-design tools move from benchmark papers into R&D pipelines, the highest-leverage integration point may not be generating final drug candidates in silico—it may be reshaping the starting genotypes that high-throughput evolution campaigns actually explore. Stability first, function second, specificity as the scoreboard.

Sources

• Nature — AI-redesigned starting points and outcomes enhance protein evolution (July 24, 2026)
• Broad Institute of MIT and Harvard — 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)
• GitHub — Nicholas-Krasnow/sequence-design-guide (ProteinMPNN redesign protocol referenced in Krasnow et al.)