Turning Prime Editing Into a Computable Reaction: OptiPrime Helps Correct KIF1A Mutations in the Mouse Brain

The challenge of precisely editing DNA is often not whether the desired sequence can be written, but how to identify a truly effective design among numerous options. A research team reported OptiPrime in *Nature Biotechnology*, breaking the prime-editing process down into a series of estimable molecular reactions and then using machine learning to predict editing efficiency. The method has progressed from cell experiments to a mouse disease model, correcting more than 40% of pathogenic KIF1A mutations in the cerebral cortex.

Prime editing uses an engineered Cas protein, reverse transcriptase, and a prime-editing guide RNA (pegRNA) to copy a specified sequence directly into the genome, without creating DNA double-strand breaks as conventional CRISPR does. However, steps including editor binding, reverse transcription, DNA strand integration, and mismatch repair influence one another, meaning that the efficiency of the same editing strategy can change substantially when applied to different sequences or cells.

OptiPrime’s distinguishing feature is that it incorporates these biological steps into the model’s structure, instead of merely searching for statistical associations in large volumes of sequence data. According to the research preprint and related patents, the model estimates reaction rates and simulates how editing outcomes change over time, covering standard prime editing, PE3—which uses an additional nick to improve efficiency—and twinPE, which uses two guide systems to carry out larger-scale rewriting. The model can also propose synonymous variants that do not alter the protein sequence, thereby reducing the likelihood that the mismatch repair system will remove the newly written DNA.

The researchers conducted prospective validation in therapeutic settings not included in the training data, including targeting CFTR p.F508del and COL7A1 p.R185X in primary human and mouse cells, as well as modifying the IL2 receptor in primary T cells. They also tested the insertion of a near gene-scale CFTR sequence using twinPE. These cases spanned point mutations, short-sequence corrections, and large DNA insertions, indicating that the model was not tuned solely for a single experimental format.

The step closest to in vivo treatment involved correcting a pathogenic mutation associated with KIF1A neurological disease. Using two AAV9 vectors, the team delivered a split PE6b editor, an optimized epegRNA, and a nicking guide RNA into neonatal heterozygous mice. Testing four weeks later showed that the correction rate at the target sequence in the cerebral cortex exceeded 40%. This demonstrates that designs selected by the model can bridge the gap between cell culture and animal tissue, but the available information is insufficient to determine whether neurological symptoms improved, and it cannot establish that the same efficiency can be reproduced in adult animals or humans.

What OptiPrime primarily shortens is the design and screening process, not the remainder of the gene therapy development pathway. Dual-AAV delivery increases the complexity of manufacturing and dose control, while unintended editing, bystander variants, immune responses, distribution across different brain regions, and long-term performance still require systematic evaluation. The model’s ability to generalize to rare sequence contexts, other tissues, and different editor versions also requires independent validation. Related methods are also covered by patents assigned to the Broad Institute and Harvard University. Whether OptiPrime can become a widely adopted design tool will depend not only on predictive accuracy, but also on reproducibility, terms of access, and regulators’ acceptance of evidence from model-assisted design.