AI-designed protein redesigns are shown to reshape the rules of evolution in a new Nature study, offering a “boosted starting point” for rapidly finding high-function enzyme variants. The work uses PACE (phage-assisted continuous evolution) to evolve a botulinum neurotoxin protease (BoNT/E) under controlled selection pressures, while testing how different AI redesigns change which mutations become beneficial.
The researchers first highlight a key challenge: stabilizing protein redesigns often reduce catalytic performance. They therefore ask whether a highly stabilized—but kinetically impaired—redesign can still outperform the wild-type (WT) enzyme during evolution. Their starting test case is ProteinMPNN redesign D4, which initially expresses well but cleaves substrate far more slowly than WT (about 20-fold lower rate), even though it shows a higher melting temperature.
To extend the comparison beyond a single redesign, the team also selects PROSS1, a top-performing alternative redesign designed to improve both stability and catalysis. Parallel PACE campaigns are then launched from D4 and PROSS1, using the same substrate panel and the same phage flow-rate schedule previously applied to WT and another redesign (D3), enabling a like-for-like assessment of how each redesigned “starting point” alters evolutionary trajectories.
Across 44 independent evolutions on multiple devices and replicates, both D4- and PROSS1-initiated populations clear selection on three substrates spanning low, intermediate, and high difficulty. Sequencing reveals that the consensus genotypes that emerge are consistently distinct from those evolved from WT or D3 backgrounds, indicating that redesign changes the fitness landscape enough to redirect evolution.
A key outcome is that the redesign-derived proteases remain more active than WT-evolved variants across all substrates. D3-evolved and PROSS1-evolved enzymes achieve similar performance, while D4-evolved variants match or exceed the other redesigns despite their slower starting catalysis—evidence that redesign can increase evolvability even when the initial enzyme is functionally handicapped.
The authors further perform a comprehensive “grafting” experiment: evolved genotypes are tested in each alternative protease background. Redesign-to-redesign compatibility is largely high—many variants remain functional when transplanted—yet D4 shows notable variability, with some genotypes becoming highly active only within their original background. They suggest this reflects a trade-off between robustness to destabilizing mutations and sensitivity to slow-catalysis-associated changes.
In contrast, grafting redesign-evolved mutations into the WT protease yields dramatically lower activity, including near-zero function on the easiest substrate and functional variants only for a minority of cases on harder substrates. Activity correlations between redesign backgrounds and WT are poor, implying that redesign enables new evolutionary solutions that WT cannot access efficiently.
Finally, the team maps genotype activity onto an embedding space from a protein language model (ESM-C). In WT, the fitness landscape is sparse: high-function solutions cluster far from the starting sequence, while local single-point mutants show little to no activity. Redesign expands the landscape—raising average fitness and allowing function in sequences closer to the starting point, including single-mutant neighborhoods—consistent with the idea that redesign accelerates adaptation by expanding the set of viable paths.
Subject of Research: AI-redesigned starting points and their effect on protein evolvability and fitness landscapes during PACE evolution
Article Title: AI-redesigned starting points and outcomes enhance protein evolution
Article References: Krasnow, N.A., Xu, J.A., Zhang, E. et al. AI-redesigned starting points and outcomes enhance protein evolution. Nature (2026). https://doi.org/10.1038/s41586-026-10820-0
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s41586-026-10820-0
Keywords: Protein engineering; AI redesign; PACE; fitness landscape; protease evolution; language model embeddings
