Cytosine base editor 4 but not adenine base editor generates off-target mutations in mouse embryos

The study addresses whether commonly used CRISPR-derived base editors differ in genome-wide fidelity when applied in mammalian embryos. Given that most human disease-associated variants are single-nucleotide changes, base editors are attractive tools for precise editing. Prior reports suggested high on-target precision but raised concerns about off-target editing, especially with cytosine base editors. This work directly compares cytosine base editor 4 (BE4) and adenine base editor (ABE) in mouse embryos, using an identical sgRNA and a trio-based whole-genome sequencing design, to quantify de novo single-nucleotide variants (SNVs) and indels and to assess proximal unintended edits at target loci.

Previous studies established base editing as a means to convert C•G to T•A (cytidine deaminase-based CBE) or A•T to G•C (adenosine deaminase-based ABE) without creating double-strand breaks. Multiple reports indicated that ABEs generally exhibit high fidelity, though unexpected C-to-T conversions within the editing window have been observed in some contexts. Whole-genome sequencing in plants (rice) and limited embryo studies reported that BE3 (an earlier CBE) can induce numerous off-target mutations, while BE4 was developed to improve specificity via an additional UGI and an optimized linker. However, comprehensive trio-based WGS comparisons in mammals remained scarce. The literature also raised concerns about potential sgRNA-independent edits, RNA off-target editing, and the need to discriminate true de novo variants from background or population variation.

• Study design: Trio-based whole-genome sequencing (WGS) comparing progeny edited with BE4 or ABE to their unedited parents and to non-injected control progeny. The same sgRNA targeting the mouse Wap gene (left sgRNA sequence: GCGACATAGCTTCCTTCTC) was used for both BE4 and ABE to control for guide-dependent effects.
• Embryo editing: In vitro fertilized C57BL/6 zygotes were cytoplasmically microinjected with in vitro transcribed mRNA encoding BE4 or ABE along with sgRNA (20 ng/µl). Embryos reaching the 2-cell stage were implanted into pseudopregnant foster mothers. Pups were genotyped by PCR and Sanger sequencing to identify on-target edits.
• Cohorts: Founders carrying edits were identified; overall, 33 mutant alleles edited by ABE and 17 mutant alleles edited by BE4 were analyzed. Group sizes used for statistical analyses included 13 non-injected controls, 13 ABE-edited, and 9 BE4-edited mice.
• DNA extraction and genotyping: Genomic DNA from tail biopsies of 3–4-week-old founders was prepared (Wizard Genomic DNA Purification Kit), PCR-amplified around the target site, and Sanger sequenced to characterize on-target edits, bystander edits, and proximal indels/deletions.
• Whole-genome sequencing and variant calling: Libraries were prepared and sequenced to high depth following institutional protocols. Reads were aligned to the mm10 reference (BWA-MEM). BAMs were processed with SAMtools and Picard to mark duplicates. Base quality recalibration and joint genotyping were performed per GATK Best Practices (GATK 3.8/4.0). Quality filtering employed BBNorm2 and GATK recommended hard filters. SNVs were filtered using criteria such as QD, FS, MQ, MQRankSum, ReadPosRankSum, and proximity to indels; dbSNP and other databases were used to exclude known variants. Indels were identified via GATK joint genotyping with recommended filters (QD, FS, ReadPosRankSum, SOR) and further curated to remove complex/repetitive overlaps and low-quality calls. Structural/complex indels were detected with LUMPY/SVTyper using discordant and split reads, applying genotype/quality thresholds and excluding repetitive element overlaps.
• Off-target prediction: Potential off-target sites were predicted with CRISPOR using mismatch and genomic context criteria; overlaps between predicted sites and observed de novo variants were assessed.
• Targeted deep sequencing: PCR amplicons spanning the target region were sequenced on Illumina platforms to sensitively detect proximal unintended base substitutions and indels.
• Statistics: De novo SNV and indel counts per animal were compared among groups using Kruskal–Wallis tests with pairwise Wilcoxon rank-sum (Mann–Whitney) tests in R (version 3.3.3). Results are summarized as mean ± SD with boxplots indicating distribution parameters.

• BE4 induces significantly more de novo single-nucleotide variants and deletions than ABE and non-injected controls across the genome in edited mouse embryos, as determined by trio-based WGS and nonparametric statistical tests (Kruskal–Wallis and Wilcoxon rank-sum).
• Proximal, unintended on-target-region mutations: BE4 caused unanticipated nearby base substitutions and deletions in 4 of 9 BE4-edited founders, whereas none of the 13 ABE-edited founders exhibited such proximal off-target mutations using the same sgRNA.
• ABE demonstrated high fidelity both at target sites and genome-wide, consistent with prior reports, although rare aberrant C-to-T conversions within the ABE editing window were noted in the literature and remain a concern.
• The majority of de novo SNVs did not coincide with in silico predicted off-target sites, suggesting a substantial component of sgRNA-independent, editor-driven mutagenesis for BE4.
• Cohort sizes for statistical analyses: 13 non-injected control mice, 13 ABE-edited mice, and 9 BE4-edited mice; on-target allele analyses encompassed 33 ABE-edited alleles and 17 BE4-edited alleles.

Using a controlled trio-based WGS design and a shared sgRNA, the study directly addresses whether base editor type influences genome-wide fidelity in mammalian embryos. The results show that BE4, despite design improvements over BE3, still introduces an elevated burden of de novo SNVs and proximal deletions compared to ABE and controls. This indicates that the cytidine deaminase-based mechanism can induce sgRNA-independent off-target DNA changes, potentially through promiscuous deamination and processing of uracil bases. In contrast, ABE maintained high fidelity, reinforcing its suitability for precise base editing applications. However, the occurrence of rare C-to-T bystander conversions within the ABE editing window and reports of RNA off-target editing underscore that even high-fidelity editors require careful validation. The findings are significant for preclinical genome editing, particularly where unintended nearby coding changes could alter protein function, and highlight the need for single-clone or single-animal analyses rather than pooled-population averages that can mask rare but consequential events.

BE4, but not ABE, generated an excess of de novo SNVs and proximal deletions in mouse embryos under otherwise matched conditions, demonstrating lower genome-wide fidelity for the cytosine editor. ABE exhibited high fidelity at target loci and across the genome, making it the preferred editor for applications requiring precision. The study recommends further optimization of cytosine base editors to reduce off-target effects and vigilant assessment of rare aberrant edits with ABE, including potential RNA off-target activity and unexpected C-to-T conversions in the editing window. Future work should expand to additional loci, sgRNAs, and editor variants, incorporate orthogonal detection methods for low-frequency mosaic mutations, and evaluate long-term phenotypic impacts.

• Variation among individual mice may influence de novo mutation counts, and group sizes, while reasonable, are modest (13 ABE-edited, 9 BE4-edited, 13 controls).
• Only a single sgRNA/target locus was used for the head-to-head comparison, limiting generalizability across genomic contexts and guide sequences.
• Trio-based WGS may miss very low-frequency mosaic variants; targeted deep sequencing mitigates but does not eliminate this limitation.
• In silico off-target predictions did not explain most de novo events, but this does not fully exclude undetected off-target binding or context-specific mutagenesis.
• Data availability is on request, limiting independent reanalysis of raw data and pipelines.