By stabilizing enzyme starting points before evolution, researchers opened new mutational routes for engineering highly specific proteases, including candidates aimed at disease-linked protein targets.
Study: AI-redesigned starting points and outcomes enhance protein evolution. Image Credit: Corona Borealis Studio / Shutterstock
In a recent study published in the journal Nature, researchers established a practical workflow, in botulinum neurotoxin (BoNT) protease models, in which artificial intelligence (AI)-based protein sequence redesign improved starting points and outcomes for automated directed evolution while helping to mitigate the stability-activity trade-offs that can constrain enzyme engineering.
The study leveraged a deep-learning protein sequence-design model called “ProteinMPNN” alongside the computational protein-stabilization method named “PROSS” to generate stabilized starting variants of botulinum neurotoxin (BoNT) proteases before subjecting them to phage-assisted continuous evolution (PACE).
Notably, the study found that in matched BoNT/E evolution campaigns, redesigned starting points yielded superior outcomes to wild-type (WT) enzymes, adapting faster and accessing a highly active mutational space that proved non-functional in the WT BoNT/E background.
Crucially, when evolved to cleave human ataxin-2, a protein implicated in neurodegeneration, an AI-redesigned protease variant achieved more than 79-fold greater specificity for the selected ataxin-2 substrate compared to the top WT-evolved enzyme, demonstrating 16% sequence divergence from the natural protein framework and highlighting the benefits of AI-assisted protein engineering and potential therapeutic-enzyme development.
Background
Since the advent of selective breeding thousands of years ago, humans have aimed to mold the biosphere to suit their needs. The directed evolution of natural enzymes is a relatively recent laboratory extension of this idea that has already helped provide feasible solutions for biomedical, therapeutic, and environmental applications.
However, as the search for novel, more efficient enzymes continues, a growing body of research emphasizes that conventional evolutionary campaigns can face a critical limitation. These campaigns are often based on screening wild-type proteins that may be constrained by marginal biophysical stability. Because many mutations conferring novel catalytic functions or non-native substrate specificities can be thermodynamically destabilizing, evolving proteins frequently lose structural integrity before reaching optimal activity. This has prompted interest in using AI tools to support the engineering of therapeutic proteins.
Computational tools like ProteinMPNN can generate protein sequences predicted to preserve a desired structure while altering much of the amino-acid sequence, but whether AI-redesigned enzymes possess superior evolvability across diverse selection pressures and complex fitness landscapes has remained unestablished.
About the study
The present study aimed to address this knowledge gap by integrating AI sequence redesign with high-throughput continuous evolution platforms. First, the study leveraged the ProteinMPNN and PROSS tools to redesign BoNT/E, BoNT/F, and BoNT/X catalytic domains.
These redesigned catalytic domains were designed to incorporate structural distance constraints (10–18 Å from the substrate and catalytic zinc ions) and multiple-sequence alignment (MSA) conservation thresholds (30–60%).
The evolutionary potential of these redesigned enzymes was evaluated using 44 parallel continuous evolution campaigns on an automated eVOLVER platform. Herein, wild-type and redesigned BoNT/E proteases were challenged against a panel of altered SNAP25 substrates of increasing difficulty (specifically, substrates 415, 413, and 412).
Finally, the workflow was applied to reprogram BoNT/E specificity toward human ataxin-2 (residues 1181–1201), a protein implicated in neurodegeneration, including ALS.
Study findings
The study's initial characterization of 74 ProteinMPNN BoNT/E designs revealed that 78% retained catalytic activity. Notably, the top-performing redesigned variants (D1-D3) were observed to demonstrate catalytic efficiencies 1.7 to 2.8 times those of the wild-type enzyme. Specifically, D2 achieved the highest catalytic efficiency (kcat/KM), up to 310 mM−¹s−¹ compared to 110 mM−¹s−¹ for WT BoNT/E.
The redesigned enzymes also demonstrated significantly superior thermal stability, with melting temperatures reaching up to 59.5°C. Notably, these numerical benefits translated well into experimental practice. The study showed that combining ProteinMPNN redesigns with mutations from a previously PACE-evolved PTEN-cleaving BoNT/E protease increased HEK293T cell expression by over 24-fold for D2 and D3, while PTEN cleavage products increased by 4.5-fold and 3.9-fold, respectively.
Furthermore, in side-by-side evolution campaigns that directly tested the effects of redesigned enzymes versus their wild-type counterparts, the redesigned starting points consistently yielded superior outcomes. Most importantly, when reviewing the outcomes of experiments conducted against the most challenging substrate, 412, WT evolutions failed in 50% of lagoons, whereas all redesigned lagoons succeeded.
The redesigned proteases were also found to tolerate destabilizing mutations (e.g., K225E) that conferred high catalytic function but caused no detectable activity when grafted into the WT background.
Furthermore, a kinetically impaired redesign (D4; 20-fold slower starting rate than WT) evolved a higher final activity than WT, supporting the view that starting-point stability can expand evolutionary potential.
Finally, the study’s ataxin-2 evolution campaigns revealed that even the top D3-evolved protease [D3(428)2] at the highest concentration tested (50 µM) displayed no detectable cleavage of the native substrate SNAP25 in a FRET assay.
Overview for integrating computational protein sequence design and continuous evolution, and redesign of stabilized BoNT/E proteases. a, Left, redesigning laboratory-evolved enzymes yields highly potent enzymes with reprogrammed specificity. Right, redesigned proteases are superior evolutionary starting points that have an expanded, higher-fitness mutation space. b, Residues that were constrained or allowed to vary during redesign in BoNT/E protease. c, Catalytic rate and soluble yield for assayed designs. Rate assays were conducted with 2.5 nM protease and 4.75 µM SNAP25 FRET substrate. The dashed lines indicate values for the WT protease. d, Abundance of top redesigned proteases in purification elution fractions E1–E4 from the soluble fraction of E. coli expression cells. Out of approximately 1 ml of elution fraction, 5 µl was loaded into each lane. e, Protein thermal melting curves to measure Tm. Thermal unfolding was followed by SYPRO orange fluorophore. Tm values, calculated as the temperature at which the slope of the curve is maximal, are provided in Supplementary Table 6. f, Michaelis–Menten plot for in vitro cleavage kinetics with the SNAP25 FRET substrate. Resulting kinetic parameters for BoNT/E protease variants are provided in Supplementary Table 5. V0, initial velocity. g, Apparent protease activity in PACE protease cleavage selection circuit host cells measured by luciferase signal. OD, optical density. The values and error bars represent mean ± s.d. of three replicates (e–g).
Conclusions
This study establishes that coupling AI sequence redesign with automated continuous evolution can, in BoNT protease models, help mitigate the classic trade-off between enzyme stability and the acquisition of new catalytic function. By expanding accessible mutational space, redesigned starting points enabled the evolution of highly specific, non-native catalytic activities.
Future studies must verify whether these advantages extend across unrelated enzyme families, thereby potentially facilitating a highly scalable framework for engineering customized therapeutic enzymes while also establishing their delivery, efficacy, and safety in disease models.
-590_16ef4f428fdb435c8c706101295dad55-80x70.jpg)
Walking protein structure explains how cells activate molecular motors