AI-designed RNA base editor alleviates mouse brain mutations and behavioral abnormalities
Background
DNA base editing permanently alters the genome, offering long-lasting therapeutic effects, but it is difficult to reverse if unexpected mutations occur. In contrast, RNA editing modifies short-lived messenger RNA (mRNA), allowing for greater control over therapeutic effects. However, C-to-U RNA editing, which converts cytosine to uracil, has been slow to translate to in vivo applications due to the enzyme's narrow sequence preferences and off-target editing.
Existing CRISPR-Cas13 approaches require the formation of a double-stranded complex between the guide RNA and the target RNA, while the cytidine deaminase APOBEC primarily acts on single-stranded RNA. Evolutionarily engineered ADAR2-based editors also raise concerns about causing A-to-I conversions in addition to C-to-U conversions. Previous animal studies achieved C-to-U editing rates of only about 20%, without significant phenotypic improvements. The researchers adopted a strategy of combining an engineered APOBEC with a PUF protein, which recognizes short RNA sequences without a guide RNA.
Key Findings
The researchers analyzed the structures of human and rodent APOBECs using AlphaFold2 and compared the phylogenetic relationships of 1,160 eukaryotic AID/APOBEC genes. They then designed a 'ProAPOBEC' library by combining multiple species' deaminase catalytic domains at both ends of human APOBEC3A. The CU-REWIRE5 series, fused with a modified PUF that recognizes 10 nucleotides, performed C-to-U editing not only in the UC context preferred by existing enzymes but also in GC, CC, and AC contexts.
In HEK293T cells, CU5.21, derived from the giant panda APOBEC, converted C832 to U in the Pcsk9 reporter, creating a stop codon, with an editing rate of 96%. This was significantly higher than the 37% of the previous CU4.1 and the CURE and xCBE based on Cas13. When CU5.21 was delivered via AAV8 to mice at a dose of 2 × 10¹² vector genomes per mouse, Pcsk9 mRNA in the liver decreased by more than 70% compared to the control group after 4 weeks. Plasma PCSK9 and cholesterol levels also decreased, but about 400 C-to-U editing events were observed throughout the transcriptome.
In the MEF2C L35P autism spectrum disorder model, AAV-PHP.eB and CU5.15, which cross the blood-brain barrier, were used. The ratio of corrected U104 in the prefrontal cortex and hippocampus was 61.3% and 68%, respectively, and the overall editing rate in several brain regions was 30-50%. MEF2C protein levels were restored, and abnormal behaviors in the social novelty preference and intruder tests were also alleviated. No bystander editing was detected near the target site C65-C150, and no C-to-T conversions were observed at the corresponding genomic locus. However, about 800 off-target C-to-U events were detected in the brain transcriptome. The original article will be published in Nature Communications in 2025.
Significance and Outlook
This platform is significant because it corrects disease-related RNA in the liver and brain without permanent DNA modification and improves behavioral phenotypes beyond molecular markers. The PUF and ProAPOBEC can be packaged in a single AAV, making the delivery construct simpler than some DNA editing methods that require two vectors. In cell experiments with other disease targets, such as APOE4, SOD1, and rhodopsin, correction rates of 85-90% were observed.
There is still a long way to go before clinical translation. High on-target editing activity and a tendency for hundreds of off-target editing events in the transcriptome have been confirmed, and if AAV expression is sustained, it may weaken the advantage of RNA editing, which is its transient nature. Immune responses, long-term toxicity, and delivery efficiency in the human brain also need to be verified separately. The small sample size of the animal study, the lack of pre-specified sample size calculations, and the lack of blinding for some evaluators are also limitations of the study design.
This input DOI is not for a new study but for an author correction published on August 12, 2026. According to the correction notice, a copy-and-paste error in Figure 7j and the source data caused the data for mouse 12 to overwrite the data for mouse 6, and the HTML and PDF have now been corrected. This panel shows the results of the social intruder test, so the corrected data should be used as the basis for interpreting the evidence of behavioral improvement. Correction notice
Nature, Published online: 12 August 2026; doi:10.1038/s41467-026-76669-zAuthor Correction: Effective in vivo RNA base editing via engineered cytidine deaminase APOBECs fused with PUF proteins
The most immediate application candidate is a therapy that lowers cholesterol by inserting a stop codon into PCSK9 mRNA in the liver. It is possible to deliver the editor to patients in a vector that allows for dose adjustment and to track plasma PCSK9, LDL cholesterol, and off-target editing in the transcriptome. In the brain, it could be a candidate for personalized treatment of rare neurodevelopmental disorders caused by single-base mutations, such as MEF2C.
Industrially, it may be possible to target mutations that are not accessible to existing editors because the target base is not UC, by using GC, CC, and AC-specific ProAPOBEC. However, at the product development stage, it is necessary to first select enzymes for each target, develop non-viral short-term delivery vectors, and establish safety criteria for the entire transcriptome. Independent replication of the corrected behavioral data is a key item to be confirmed before clinical development.