PE8 with an AI-Redesigned Reverse Transcriptase Improves Prime Editing Efficiency in Human Primary Cells and Mice
Improved editing efficiency and protein stability
A study published in Nature Biotechnology reports that redesigning reverse transcriptase (RT) with artificial intelligence (AI) improved the protein stability and editing efficiency of a prime editor. PE8, developed by the researchers, showed improved editing efficiency in vitro, including in human primary cells, and editing efficiency up to 2.9 times that of existing comparator editors in mice.
Protein engineering and laboratory evolution have been used to improve prime editor performance. However, mutations that increase editing efficiency do not necessarily improve other protein properties. In this study, the researchers found that existing efficiency-enhancing mutations impaired protein stability and expression levels, limiting performance. They therefore applied a redesign strategy to improve RT stability and expression while preserving regions required for catalytic activity.
RT redesign increased protein levels and editing efficiency
The researchers used structure-informed AI methods, including the inverse-folding neural network ProteinMPNN, to redesign the RT region of prime editors improved by protein engineering and laboratory evolution. They report that regions essential for catalytic activity were preserved. The redesigned RTs contained 30–163 amino acid substitutions and showed improved folding stability and soluble expression.
After messenger RNA (mRNA) delivery, intracellular prime editor protein levels were up to twice those of the comparators. This figure compares editor protein levels inside cells and was reported separately from improvements in folding stability and soluble expression. It should not be interpreted as a doubling of soluble expression itself or as evidence of a longer protein half-life.
The redesigned PE8 improved editing efficiency under in vitro conditions involving several human primary cell types and multiple delivery methods. In mice, editing efficiency was up to 2.9 times that obtained with PE6, PE7, and PEmax. The measured outcome was editing efficiency; it does not indicate faster editing over time.
The provided abstract does not give figures by cell type or delivery method, identify the target tissues in mice, or report absolute editing efficiencies and comparator-specific results. The relative maximum of 2.9 times therefore cannot be generalized to every target or delivery condition.
Structure-based design complements laboratory evolution
This study provides evidence that structure-based redesign can further improve editors developed through laboratory evolution. The researchers describe this as a generalizable approach to improving genome editors that complements laboratory evolution. Its significance lies in a design strategy that considers protein stability and expression levels alongside editing efficiency.
The results in human primary cells and mice are preclinical evidence. They do not establish therapeutic benefit or clinical safety in patients, and whether the same improvements apply to other genome editors requires separate validation. To assess the potential scope of application, the targets, delivery conditions, sample sizes, statistical uncertainty, and safety assessments need to be checked in the full paper and supplementary materials.
The study description is based on the provided Europe PMC bibliographic record and abstract.
Nature Biotechnology, Published online: 21 May 2026. DOI: 10.1038/s41587-026-03149-6Summary: While computational evolution has optimized the catalytic parameters of therapeutic genome editors, select lineage mutations systematically compromise baseline protein stability and expression kinetics to restrict overall in vivo performance. To bypass this thermodynamic bottleneck, this study implements a structure-informed, artificial intelligence-guided framework employing the inverse-folding neural network ProteinMPNN to systematically redesign the reverse transcriptase (RT) domains of prime editors. By executing 30 to 163 programmatic amino acid substitutions while preserving active catalytic domains, the engineered PE8 variants achieved over a twofold enhancement in soluble expression profiles and intracellular protein availability following mRNA delivery. Deployed across divergent ex vivo human primary cellular matrices and in vivo murine targets, the PE8 architecture demonstrated an up to 2.9-fold elevation in gene alteration velocity compared to state-of-the-art PE6, PE7, and PEmax configurations, establishing a highly scalable baseline for programmable enzyme stabilization.
Teams developing prime editors can assess RT stability and expression levels alongside editing efficiency when evaluating candidates with improved editing efficiency. For example, they could redesign a previously improved RT using structure-based AI, then compare protein levels and editing efficiency in the intended human primary cells and under the planned delivery conditions. The PE8 results provide preclinical evidence supporting this candidate optimization strategy. However, protein levels of up to twice the comparator levels and editing efficiency of up to 2.9 times the comparator efficiency are relative values reported under their respective conditions. They cannot be translated into therapeutic benefit for patients or reduced dosing; performance and safety must be validated separately under the intended conditions before a candidate is selected for development.