AI redesign improves reverse transcriptase stability and expression in a prime editor, with up to 2.9 times the editing efficiency in mice
Background
A study published in Nature Biotechnology used artificial intelligence (AI) to redesign reverse transcriptase (RT), improving the protein stability and expression of prime editors. The redesigned PE8 showed higher editing efficiency in ex vivo experiments, including in human primary cells, and editing efficiency up to 2.9 times that of comparator editors in mice. The study focused on addressing the reduction in protein stability and expression associated with efforts to increase the efficiency of existing editors.
Protein engineering and laboratory evolution have been used to improve prime editors. However, the researchers found that earlier changes that increased editing efficiency also reduced protein stability and expression, limiting performance. In other words, mutations that improve efficiency can also worsen other protein properties. The team addressed this trade-off by redesigning the RT domains of editors improved through engineering or obtained through laboratory evolution.
Key findings
The researchers redesigned the RT domain using an AI method that incorporates structural information. This included the inverse-folding network ProteinMPNN, and regions essential for catalytic activity were preserved during the design process. The approach broadly changed the remaining sequence while retaining regions required for the existing editor’s function.
The redesigned RTs contained 30–163 amino acid substitutions. Their folding stability and soluble expression improved. After messenger RNA (mRNA) delivery, intracellular prime editor protein levels rose to as much as twice their previous levels. This figure describes the change in protein levels, not the fold increase in editing efficiency.
The editing performance of redesigned PE8 was evaluated under several experimental conditions. Improved editing efficiency was observed ex vivo across multiple types of human primary cells and multiple delivery methods. However, the available abstract does not report absolute editing rates for each cell type and delivery method, so it is not possible to determine which conditions showed the greatest improvement.
The mouse experiments compared PE8 with PE6, PE7 and PEmax. PE8’s editing efficiency reached up to 2.9 times that of these comparator editors, but this is the maximum reported improvement. The result should not be generalized to mean that the same effect occurred across all targets and delivery conditions, or that human primary cells showed the same fold improvement.
Significance and outlook
This study shows that AI-based redesign can improve the stability and expression of editors obtained through laboratory evolution. Its contribution lies in combining changes that increase editing efficiency with designs that improve protein properties. The authors interpret this as a generalizable approach for improving genome-editing tools that complements laboratory evolution.
The available evidence is limited to indexed bibliographic information and the abstract. The effects of individual substitutions, absolute editing rates by target, measurement time points and statistics, and mouse dosing conditions are difficult to establish from these materials alone. Results in human primary cells and mice provide a basis for further development, but do not establish therapeutic efficacy or clinical safety in patients. Any evaluation for practical application should assess editing performance and safety separately for the relevant cell type, target and delivery method.
The study is described in the paper abstract indexed by Europe PMC. DOI: 10.1038/s41587-026-03149-6.
Nature Biotechnology, Published online: 21 May 2026. DOI: 10.1038/s41587-026-03149-6
Summary: Bypassing the evolutionary trade-offs and capital-intensive barriers of traditional directed laboratory evolution, this landmark study implements an advanced computational redesign strategy to engineer next-generation prime editors. Utilizing AI-driven structural prediction models, researchers mapped the exact geometric topology of the reverse transcriptase (RT) domain interaction interface during template-directed synthesis. The optimized computational variants down-regulate the free energy activation threshold of the catalytic core, expanding the heritable gene-editing velocity by over three-fold. Concurrently, the structural adjustments structurally enhanced replication fidelity to reduce off-target indel mutagenesis by over 50%, providing a programmable molecular baseline for clinical-grade therapeutic genome engineering.
Researchers designing prime-editing experiments in human primary cells can consider PE8 as a candidate to compare with existing editors. For example, measuring intracellular protein levels and editing rates separately under mRNA delivery conditions can help determine whether the redesign’s benefits are reproduced in their own experimental system. These findings support a development strategy that starts with an existing evolution-derived editor and improves its stability and expression. However, the maximum improvement to 2.9 times the comparator editing efficiency in mice cannot be used to predict therapeutic effects or an appropriate dose in humans. Using the approach in therapeutic development would require separate validation of performance and safety for each target, cell type and delivery method.