For decades, researchers have used laboratory evolution to create proteins with new and useful functions. The process can be remarkably effective, but it comes with a major limitation: as proteins evolve new capabilities, they often lose stability. That instability can reduce performance and limit how far scientists can push protein engineering for therapeutic applications.
A team led by David Liu at the Broad Institute discovered a powerful solution by combining artificial intelligence with laboratory evolution. The researchers first used the AI model ProteinMPNN (a deep learning model for predicting amino acid sequences for protein backbones) to redesign naturally occurring proteins, creating versions predicted to be far more stable while maintaining their overall structure. They then applied phage-assisted continuous evolution (PACE), a technology pioneered in the Liu laboratory, to evolve those proteins toward new functions.
The strategy produced proteins that consistently outperformed those evolved from natural starting points. In one demonstration, the team engineered a botulinum neurotoxin protease to cut the disease-associated protein ataxin-2, achieving activity that was 79 times greater than versions evolved from the natural enzyme. The resulting proteins were also more stable and more specific.
The findings suggest that AI-designed proteins can serve as superior starting materials for protein engineering, opening new possibilities for developing therapeutic enzymes and molecular tools. Research computing was central to the breakthrough, enabling AI-driven protein design, large-scale sequence analysis, and the computational evaluation of protein variants that guided subsequent evolution experiments.
Krasnow N.A., Xu J.A., et al. (2026), AI-redesigned starting points and outcomes enhance protein evolution, Nature, doi: 10.1038/s41586-026-10820-0
