Optimizing CRISPR precision in mouse embryos via microhomology-mediated end joining-dominant targeting

Abstract

CRISPR/Cas9 technology enables efficient gene editing in mice, but its reliance on non-homologous end joining often leads to unpredictable and mosaic mutations in founder (F0) animals. Here, we present a hybrid genome editing strategy that combines in silico prediction software with in vitro validation using mouse embryonic stem cells (mESCs). Although the software was trained on mESC datasets, actual editing outcomes in mESCs more accurately reflected mutation patterns observed in blastocysts and post-implantation embryos. Using this information to develop an integrated pipeline, we pre-selected guide RNAs (gRNAs) predicted to promote microhomology-mediated end joining (MMEJ)-dominant repair and validated them in mESCs prior to embryo injection. Applied to the Tyr and Fgf10 genes, this approach enabled efficient generation of F0 mice with highly uniform genotypes. Our strategy enhances the predictability and reproducibility of CRISPR-based genome editing in mice and may help reduce animal usage in gene editing studies.

A hybrid CRISPR strategy combining in silico prediction of microhomology-mediated repair with validation in mouse embryonic stem cells enables predictable genome editing and efficient generation of genetically homogeneous founder mice.

Introduction

Genetically modified animals play a pivotal role in modeling human diseases and advancing therapeutic development. Various species, including rodents, such as mice and rats, as well as larger animals, such as pigs and non-human primates, are used for the in vivo replication of human genetic disorders. These models provide valuable insights into the molecular and cellular mechanisms of disease onset and progression^1–4, serving as essential platforms for the preclinical evaluation of therapeutic safety, efficacy, and delivery strategies. By providing a biologically relevant context, they facilitate the development of targeted and personalized treatments^5–8. At the same time, ongoing efforts to reduce, refine, and replace the use of laboratory animals aim to strike a balance between scientific advancement and ethical responsibility^9–11.

Over the past decade, the CRISPR/Cas9 system has revolutionized gene editing, enabling the generation of knockout (KO) mice with nearly 100% efficiency in the F0 generation^12–15. Despite these advancements, a major challenge persists: researchers cannot fully control the outcomes of double-stranded DNA breaks (DSBs). These breaks are repaired by endogenous repair pathways, including the non-homologous end joining (NHEJ), microhomology-mediated end joining (MMEJ), which uses short homologous DNA sequences to align broken ends, and the more precise homology-directed repair (HDR) pathway¹⁶. HDR is highly precise but suffers from low on-target efficiency^17,18. NHEJ and MMEJ have higher efficiency¹⁹; however, NHEJ results in stochastic indel genotypes²⁰, while MMEJ introduces highly stereotyped deletions²¹.

Current approaches for generating KO animals typically prioritize guide RNA (gRNA) selection based on editing efficiency and minimization of off-target effects^22,23. Editing efficiency refers to how effectively a given gRNA introduces indels at the intended target site after DSBs. However, mutational outcomes are often unpredictable and rarely considered during gRNA design, leading to a high incidence of mosaicism in F0 animals. Recently, advances in machine learning have enabled the prediction of DNA repair outcomes by training deep learning or regression models on large sequencing datasets from cultured cells, revealing the frequency and patterns of insertions and deletions following CRISPR/Cas9-induced DSBs^21,24–28. Among these, inDelphi, trained on a large dataset of CRISPR/Cas9-induced indels in human and mouse cells, shows strong performance in non-mammalian vertebrates, including Xenopus tropicalis, Xenopus laevis, and Danio rerio embryos²⁹; however, its applicability in mammalian zygotes remains to be validated.

In this study, we integrated a machine learning model (inDelphi) favoring MMEJ into the gRNA design process. Using an in vitro mouse embryonic stem cell (mESC) culture system, we optimized and validated mutational predictions, enabling the generation of genetically uniform and predictable mutant mice. gRNAs targeting the Tyr gene were tested in both blastocysts and mESCs. Bulk genotyping by amplicon sequencing showed moderate correlations with inDelphi predictions, with stronger concordance observed in mESCs. To assess single-embryo resolution, we performed embryo transplantation and analyzed E11.5 embryos, confirming that gRNAs with higher predicted precision produced uniform genotypes. Applying this approach to the Fgf10 gene, we successfully generated the predicted mutations in F0 embryos. These findings demonstrate an efficient, predictive pipeline for generating homogeneous, mutation-specific mice in the F0 generation, enabling more precise modeling of human pathogenic mutations and advancing disease research.

Results

Generation of genetically modified embryos guided by inDelphi-predicted gRNA design

To develop a pipeline for generating predictable and homogenous mice, we targeted exon 1 of the Tyr gene in mice, whose disruption results in an albino phenotype^30,31 (Fig. 1a). Based on the presence of NGG PAM sequences, 102 candidate gRNAs were identified. From these, 14 gRNAs were randomly selected for further evaluation. Their mutational outcomes were analyzed using the mESC-trained version of inDelphi. We chose this model because mESCs, while not identical to zygotes, are presumed to share similar transcriptional features with early embryonic stages³². inDelphi provides a precision score (Supplementary Fig. 1a), defined as a statistical measure (ranging from 0 to 1) of the predicted indel frequency distribution for a given gRNA. The most frequent genotype (MFG) refers to the specific genotype that occurs most often, and its frequency provides a simple and practical measure. (Fig. 1b). In addition, inDelphi reports a microhomology (MH) strength score, which shows a strong positive correlation with both the precision score and the frequency of the MFG. The model also provides a predicted frameshift frequency, indicating the likelihood of coding sequence disruption. Off-target effects were assessed using CRISPOR³³, which provides both MIT³⁴ and CFD³⁵ specificity scores (Supplementary Fig. 1a). All selected gRNAs had MIT scores above 50 and CFD scores above 70, indicating low predicted off-target activity. These results support the high specificity and suitability of the gRNAs for downstream applications.

Fig. 1. Predictive accuracy of inDelphi for CRISPR-induced mutation patterns in mouse blastocysts.

a Workflow of the experimental procedures. b Schematic diagram of the Tyr gene in mice, showing gRNA target sites within exon 1. Corresponding inDelphi predictions are indicated for each gRNA. Blue boxes denote microhomology (MH) sequences. c The most frequent mutant genotype observed in mouse blastocysts, as determined through a CRISPResso2 analysis of NGS data. Wild-type reads were excluded from the analysis, and percentages were calculated based on the remaining mutated reads. Blue boxes denote MH sequences, and red boxes indicate insertions, orange lines indicate the PAM sequence. d Comparison of the most frequent genotype (MFG) predicted using inDelphi and the frequency of the corresponding genotype observed in the blastocyst experimental data. e Correlation between inDelphi-predicted mutation patterns and experimental results in blastocysts. Correlation strength is defined as follows: weak (r < 0.4), moderate (0.4 ≤ r < 0.7), strong (0.7 ≤ r < 0.9), and very strong (r ≥ 0.9). f Scatter plots comparing inDelphi-predicted and observed mutation frequencies in blastocysts for all gRNAs. Each point represents a distinct indel pattern. Pearson correlation coefficients indicate concordance between prediction and outcome. Log10-transformed frequencies are shown; a value of 1 was added to all predicted and observed frequencies to avoid log(0). g Stacked bar chart showing the distribution of repair outcome categories for each gRNA. Mutation outcomes were classified as MH ≥ 2 bp, MH = 1 bp, MH-less, 1 bp insertion, and others. h Violin plots showing deletion-size distributions across MH length categories (1 bp, 2–3 bp, ≥4 bp). Boxplots indicate the median and interquartile range, and each dot represents a single indel genotype. Statistical comparisons were performed using the Kruskal–Wallis test followed by pairwise Wilcoxon rank-sum tests (Benjamini–Hochberg correction). Significance levels are indicated as: * P < 0.05, ** P < 0.01, *** P < 0.001, and **** P < 0.0001. i Scatter plot showing the relationship between Shannon entropy and the inDelphi MH strength score for each gRNA. A strong negative correlation (r = −0.75) indicates that higher MH strength corresponds to more homogeneous editing outcomes. j Editing efficiency, calculated as the percentage of mutated reads after subtracting wild-type sequences. k Correlations of editing efficiencies between blastocyst experimental results and both in vitro (EGxxFP) and in silico prediction tools.

We then performed in vitro fertilization using C57BL/6Jms mice. After electroporation of the gRNA-Cas9 ribonucleoprotein (RNP) complex, fertilized eggs were cultured until the blastocyst stage (Supplementary Fig. 1b). The blastocyst development rate was approximately 40% (530 out of 1364) in CRISPR-edited mouse embryos (Supplementary Table 1). Genomic DNA was extracted, and the target sequence was amplified using primer sets producing ~300 bp amplicons for each gRNA and subjected to bulk next-generation sequencing (NGS). On average, approximately 35 blastocysts were analyzed per gRNA. NGS data were analyzed using CRISPResso2³⁶, which provides all mutation types and corresponding read percentages (Supplementary Fig. 1d). We excluded the wild-type sequences from the NGS results because they are ignored by inDelphi when calculating the MFG for each gRNA (Fig. 1c, Supplementary Fig. 1c). The MFG values predicted using inDelphi were similar to the frequencies of the corresponding genotypes in blastocysts (Fig. 1d). Correlations between all predicted and observed mutation patterns were evaluated for each gRNA (Fig. 1e, Supplementary Fig. 3). Only gRNA1 showed a very strong correlation (defined as r ≥ 0.9), while five gRNAs showed a strong correlation (0.7 ≤ r < 0.9), three showed a moderate correlation (0.4 ≤ r < 0.7), and five showed a weak correlation (r < 0.4). The overall correlation between predicted and observed mutation frequencies across all gRNAs was moderate (r = 0.64) (Fig. 1f), indicating a statistically significant positive association and supporting the potential utility of inDelphi for genome editing design in mouse embryos. In our dataset, the average predicted MFG was 30.4%, while the average observed frequency in mouse blastocysts was 32%, indicating strong agreement between predicted and experimental outcomes. The highest predicted MFG was 69.8% for gRNA1, with a corresponding experimental frequency of 60%, demonstrating that certain gRNAs can induce highly dominant mutational outcomes in vivo. These results highlight the value of inDelphi for identifying gRNAs that yield consistent and high-frequency genotypes in mouse embryos.

Next, we classified blastocyst mutation outcomes into five repair outcome categories: MH ≥ 2 bp, MH = 1 bp, MH-less, 1 bp insertion, and others (Fig. 1g). This classification enabled visualization of the relative contributions of MH-associated and non-MH-associated events across gRNAs. We then analyzed the indel size distributions to examine their relationship with MH length (Fig. 1h). Longer MH sequences were associated with larger deletions, indicating a length-dependent influence of MH on deletion size. To further quantify the diversity of mutational outcomes in blastocysts, we calculated Shannon entropy (hereafter referred to as entropy) for each gRNA based on the observed indel distributions (Fig. 1i). Entropy provides a quantitative measure of mutational diversity by incorporating both the number of distinct indel species and their relative frequencies. Higher entropy values reflect more heterogeneous repair outcome distributions, whereas lower values indicate more uniform and biased outcomes. Notably, entropy showed a strong negative correlation with the inDelphi MH strength score (r = −0.75), indicating that higher MH strength is associated with more homogeneous editing outcomes in embryos.

Next, we used the EGxxFP single-strand annealing (SSA) assay to validate gRNA-dependent cleavage activity in HEK293FT cells, following a previously established method³⁷ (Supplementary Fig. 2). Briefly, the pCAG-EGxxFP plasmid containing an insertion of exon 1 from the mouse Tyr gene and a plasmid expressing hCas9 along with a sgRNA expression cassette targeting 14 sites were co-transfected into HEK293FT cells. Cleavage activity was evaluated by comparing fluorescence intensities using flow cytometry. EGxxFP assay results are expressed as the percentage of GFP-positive cells relative to the background fluorescence of a vector lacking a gRNA, with values exceeding the baseline indicating cutting activity of the tested sgRNA. While the assay does not quantify editing efficiency, we evaluated the extent to which its fluorescence signal reflects the editing efficiencies observed in embryos. In blastocysts, the average editing efficiency was 43.2%, as calculated using CRISPResso2, which reports the proportion of reads containing mutations (excluding wild-type sequences). Editing efficiency varied across different gRNAs (Fig. 1j). We observed a moderate correlation between the EGxxFP assay and experimental results (r = 0.64), supporting the assay’s utility in predicting in vivo CRISPR activity (Fig. 1k). Additionally, we evaluated the blastocyst results using commonly employed machine learning models for editing efficiency prediction. CRISPRon showed the highest correlation (r = 0.52); however, the EGxxFP assay demonstrated good predictive performance overall.

Development of a complementary mESC-based assay to benchmark embryonic outcomes and improve inDelphi predictions

To improve predictive accuracy, we developed an in vitro assay using mESCs to evaluate both mutational outcomes and editing efficiency. mESCs were transfected with SpCas9-2A-GFP plasmids expressing the gRNA of interest. Three days after transfection, GFP-positive cells were isolated using fluorescence-activated cell sorting (FACS), and editing outcomes were analyzed using NGS of PCR-amplified gRNA target sites. Editing efficiency in mESCs was similar to that observed in blastocysts, with a Pearson correlation coefficient of 0.65, indicating a significant positive association (p < 0.05; Fig. 2a, b).

Fig. 2. Comparative analysis of mESCs and inDelphi in predicting CRISPR editing outcomes in mouse blastocysts.

a Comparison of editing efficiencies of mouse blastocysts and mESCs. b Scatter plot illustrating the relationship between editing efficiencies in blastocysts and mESCs. c Most frequent genotype (MFG) in mESCs, as identified through a CRISPResso2 analysis of NGS data. Wild-type reads were excluded, and percentages were calculated from the remaining mutated reads. Blue boxes denote microhomology (MH) sequences, and red boxes indicate insertions, orange lines indicate the PAM sequence. d Comparison of the MFG observed in mESCs and the frequency of the corresponding genotype identified in blastocysts. e Correlation analysis between mutation patterns in mESCs and blastocyst experimental results. f Scatter plots comparing mESC-derived and blastocyst mutation frequencies for all gRNAs. Each point represents a distinct indel pattern. Pearson correlation coefficients indicate the level of concordance. Log10-transformed frequencies are shown; a value of 1 was added to all predicted and observed frequencies to avoid log(0). g Comparative analysis of the correlation strength between inDelphi predictions and mESC results with blastocyst experimental data. h Box plot comparing the predictive accuracy of inDelphi and mESCs for blastocyst editing outcomes. Each point represents the correlation for a single gRNA. Box plots indicate the median and interquartile range, with whiskers showing the full data range. A significant difference between the two groups was detected (p < 0.05). i Scatter plots comparing mESC-derived and inDelphi-predicted frequencies for all gRNAs. Each point represents a distinct indel pattern. Pearson correlation coefficients indicate the level of concordance. Log10-transformed frequencies are shown; a value of 1 was added to all predicted and observed frequencies to avoid log(0).

All mutation patterns were qualified in terms of read counts and percentages (Supplementary Fig. 4). Wild-type sequences were excluded from the NGS results, and the MFG for each gRNA was evaluated (Fig. 2c). The MFGs observed in mESCs were similar to those in blastocysts (Fig. 2d). The correlation between mESC and blastocyst mutation patterns was analyzed for each gRNA (Fig. 2e, Supplementary Fig. 5). Among the tested gRNAs, five exhibited a very strong correlation, four showed a strong correlation, four displayed a moderate correlation, and one demonstrated a weak correlation. There was a strong, significant positive correlation between the summed mutation patterns across all gRNAs in mESCs (r = 0.82; Fig. 2f), closely mirroring the trends observed in the individual gRNA analyses. The mutation pattern in blastocysts using inDelphi (avg. r = 0.57 ± 0.17) and mESCs (avg. r = 0.76 ± 0.21) for each gRNA differed significantly (Fig. 2g, h). Notably, correlation between the mutation pattern in mESCs and inDelphi predictions was moderate (r = 0.72, Fig. 2i), although we used mESC-trained version of inDelphi. Our results demonstrate that incorporating the mESC culture system to assess mutation patterns and editing efficiency enhances the predictive power of inDelphi.

Post-implantation validation of predicted genome editing outcomes

Following validation, edited embryos were transplanted and dissected at E11.5 to assess retinal pigmentation and confirm genotypes using whole-body lysates analyzed by NGS (Fig. 3a). We selected gRNAs 1, 2, and 9 to represent a range of editing outcomes. gRNA1 had the highest predicted MFG (69.8%) and showed a very strong correlation between inDelphi predictions and both blastocyst and mESC mutation patterns. gRNA2 had a predicted MFG of 31% and exhibited strong correlations in both comparisons. gRNA9 had a lower predicted MFG (7.9%), and showed weaker correlations (i.e., weak between inDelphi and blastocysts and moderate between mESCs and blastocysts).

Fig. 3. Validation of inDelphi and mESC predictions at single-embryo resolution in E11.5.

a Workflow of the experimental procedures. Created in BioRender. Okamura, E. (2026) https://BioRender.com/l30xep1. b Phenotypic comparison of edited and wild-type E11.5 embryos. White arrows indicate retinal pigmentation. c Editing efficiency of E11.5 embryos; the red bar denotes the mESC editing efficiency for reference. d Mutation patterns of each embryo. Blue indicates the MFG predicted using inDelphi; green indicates wild-type. Each column represents a single embryo. e Schematic showing primer positions and the expected long-range PCR amplicon (5068 bp) for the Tyr gene. Gel electrophoresis images display PCR products from embryos initially classified as homozygous by short-amplicon NGS, including wild-type and edited embryos, alongside a molecular size marker. Each lane represents an individual embryo. f Comparative analysis of the correlation strength between inDelphi predictions and mESC results with E11.5 experimental data. g Frameshift allele percentages for each embryo, grouped by pigmentation phenotype (no pigmentation, partial pigmentation, full pigmentation) across gRNA1, gRNA2, and gRNA9. Bars represent mean ± SD, and individual embryos are plotted as points. Scale bar = 2 mm.

To validate the predictive performance of selected gRNAs at post-implantation stages, we next analyzed embryos recovered at E11.5. As an initial phenotypic readout, we assessed retinal pigmentation in E11.5 embryo images and classified embryos into three categories: no pigmentation, partial pigmentation, and full pigmentation based on magnified observations and quantified eye pigmentation intensity (Fig. 3b, Supplementary Fig. 6a). For gRNA1, five of eight embryos exhibited no pigmentation, one showed partial pigmentation, and two displayed full pigmentation. The three embryos with partial pigmentation showed comparatively lower editing efficiencies (61.6%, 79.8%, and 50.5%) (Fig. 3b, c). All five embryos edited with gRNA2 displayed the expected loss of pigmentation, although two showed relatively lower editing efficiencies (78.9% and 76.1%). For gRNA9, all eight embryos exhibited high editing efficiencies approaching 100%, among which five showed no pigmentation and three showed partial pigmentation. Overall, editing efficiencies were consistent with those observed in mESCs (Fig. 3c). Phenotypic penetrance was consistent with editing efficiency for gRNA1 and gRNA2; however, despite uniformly high editing efficiencies in all gRNA9 embryos, some exhibited partial pigmentation (Supplementary Fig. 6b). To further examine this relationship, we calculated Shannon entropy for each embryo to quantify allele diversity and assessed its correlation with editing efficiency. A very strong negative correlation was observed for gRNA1 and gRNA2 (r = −0.90 and −0.86), indicating that embryos with higher editing efficiencies tended to have more homogeneous genotypes. In contrast, a negligible correlation was detected for gRNA9 (r = 0.16), likely because all embryos exhibited nearly 100% editing efficiency (Supplementary Fig. 6c).

Genotyping analysis based on NGS data revealed that for gRNA1, MFG predicted by inDelphi—a 7 bp deletion—was predominantly observed across embryos (Fig. 3d). Three embryos (nos. 1, 7, and 8) carried only the 7 bp or 11 bp deletion allele, suggesting homozygosity. Four embryos (nos. 2, 3, 4, and 6) harbored two alleles, while one embryo (no. 5) exhibited three alleles (Fig. 3d). Because large deletions and insertions are known to occur during CRISPR editing^38–42, and our short-read NGS analysis was limited to ~300 bp PCR amplicons, we examined the possibility of large deletions in the three putative homozygous embryos (nos. 1, 7, and 8) using primers designed to amplify a ~ 5 kb genomic region (Fig. 3e). All three embryos exhibited PCR bands comparable to those of WT control. We further analyzed the PCR amplicons by nanopore-based sequencing followed by DAJIN2 analysis⁴³, a long-read genotyping pipeline (Supplementary Fig. 7a, b). All three embryos carried a uniform 7 bp deletion, indicating that they are homozygous.

For gRNA2, all embryos carried the inDelphi-predicted 7 bp deletion but displayed variable allele compositions. In contrast, for gRNA9, inDelphi predicted a 3 bp deletion; however, only two embryos harbored this mutation, whereas the remaining embryos exhibited heterogeneous genotypes. Notably, one embryo (no. 8) appeared to carry only a 26 bp deletion allele, suggesting homozygosity based on short-read NGS analysis. To further investigate this possibility, we performed long-range PCR, which revealed the presence of a ~ 0.7 kb deletion (Fig. 3e, Supplementary Fig. 7a). The detailed analysis by long-read genotyping pipeline with nanopore sequencing revealed 700 bp deletion and 26 bp deletion alleles (Fig. 3d, Supplementary Fig. 7b), indicating that embryo no.8 was in fact heterozygous⁴².

We next compared the correlations between inDelphi predictions, mESC results, and E11.5 genotypes across all detected mutation patterns (Fig. 3f). Similar trends were observed for gRNA1 and gRNA2, whereas gRNA9 showed higher agreement between mESC and in vivo data than with inDelphi predictions. To further characterize the repair profiles, embryo mutation outcomes were classified into five categories: MH ≥ 2 bp, MH = 1 bp, MH-less, 1 bp insertion, and others (Supplementary Fig. 8). As expected, gRNA1 predominantly yielded MH ≥ 2 bp outcomes, whereas gRNA2 exhibited a mixture of MH ≥ 2 bp, MH = 1 bp, and other mutational outcomes. gRNA9 displayed diverse patterns without a clear preference for MH-associated repair. These results were consistent with the inDelphi-predicted MH strength scores: gRNA1 showed a high MH strength (2.11) and produced highly homogeneous mutation profiles, gRNA2 had a lower score (0.5) corresponding to more diverse allelic outcomes, and gRNA9 exhibited largely stochastic mutation patterns associated with a negative score ( − 0.11). We also observed MMEJ-associated mutation patterns (Supplementary Fig. 9). For gRNA1, a 7 bp deletion accounted for 75.7% of mutations, followed by an 11 bp deletion at 12.2%. For gRNA2, the three most frequent MMEJ-associated mutation patterns occurred at 26.9%, 21.3%, and 13.2%, respectively. For gRNA9, the two most frequent MMEJ-associated deletions occurred at 24.2% and 14.2%. Based on these distributions, we hypothesize that the high homogeneity observed for gRNA1 arises from a dominant microhomology, whereas the increased diversity observed for gRNA2 reflects the presence of multiple competing microhomologies of comparable strength. In contrast, the balanced contribution of MMEJ- and NHEJ-mediated repair at the gRNA9 locus suggest largely stochastic mutation outcomes. In addition, based on the number of detected alleles per embryo, the timing of cleavage relative to cell division appears to differ among gRNAs: cleavage likely occurred predominantly at the 1–2 cell stage for gRNA1 and gRNA9, whereas it extended up to the 8-cell stage for gRNA2 (Supplementary Fig. 10). Finally, we calculated the proportion of frameshift alleles for each gRNA to relate genotypic outcomes to phenotypic penetrance, and found that embryos with higher frameshift frequencies exhibited higher loss-of-pigmentation phenotypes (Fig. 3g).

These findings suggest that selecting MMEJ-dominant targets improves the genetic homogeneity and predictability of mutations in post-implantation mouse embryos.

Proof of principle for an optimized gene editing pipeline in the Fgf10 gene

To validate our approach, we applied the pipeline to exon 3 of the Fgf10 gene, whose KO results in a limbless phenotype⁴⁴. inDelphi identified 15 candidate gRNAs targeting this exon. We selected the target with the highest predicted MFG—a 4 bp deletion with a frequency of 39.3%—and the highest frameshift rate among all candidates (88.5%) and MH strength of 1.21. (Fig. 4a, b). This gRNA was first used to transfect mESCs, followed by electroporation of the corresponding RNP complex into mouse embryos. The blastocyst development rates were 45.7% and 57.1% in two independent experiments (Supplementary Table 1). The MFG predicted using inDelphi was well represented in both systems, with frequencies of 41% in mESCs and 53% in blastocysts, indicating strong predictive relevance in vivo (Fig. 4c). Correlation coefficients between blastocyst mutation patterns and those predicted using inDelphi and observed in mESCs were 0.94 and 0.97, respectively (Fig. 4d). Editing efficiencies were similarly high: 94% in mESCs and 92% in blastocysts (Fig. 4e). We then transplanted electroporated embryos into surrogate females and analyzed them at embryonic day E15.5 to assess limb development and body size. All six embryos edited with this gRNA exhibited the expected limbless and reduced-body-size phenotypes, confirming the robustness of this editing strategy (Fig. 4f). Editing efficiencies in E15.5 embryos reached nearly 100%, closely matching those observed in mESCs and blastocysts (Fig. 4e, g). We identified three embryos that appeared to be homozygous and, suspecting the presence of large deletions, performed long-range PCR followed by nanopore sequencing and genetic analysis using DAJIN2⁴², which revealed large deletions in two embryos (Fig. 4h, i, Supplementary Fig. 11a, b). Although no homozygous embryos carrying the 4 bp deletion were obtained, we summarized all mutation patterns to compare embryonic outcomes with inDelphi predictions and mESC data (Fig. 4j). This analysis showed that the 4 bp deletion was the dominant mutation across all datasets and was similarly enriched in both inDelphi predictions and mESC results. In addition, the 1 bp insertion and 13 bp deletion, which were ranked as the second and third MFGs in inDelphi and mESC analyses, were also frequently observed in E15.5 embryos. Overall, mutation frequencies in E15.5 embryos (68.7%), inDelphi predictions (66.2%), and mESCs (59.6%) were distributed across the same four major mutation types, with all other unpredicted variants grouped as “Others”.

Fig. 4. Validation of the established gene-editing pipeline for the Fgf10 gene in mice.

a Workflow of the experimental procedures. b Schematic diagram of the Fgf10 gene in mice, showing gRNA target site within Exon 3, with corresponding inDelphi predictions indicated for the selected gRNA. Blue box denotes microhomology (MH) sequence. c Comparison of the most frequent genotype (MFG) predicted by inDelphi with the frequencies observed in mESC and blastocyst data. d Scatter plots comparing mutation frequencies for all gRNAs. Left: inDelphi-predicted vs. observed mutation frequencies in blastocysts. Right: mESC-derived vs. blastocyst mutation frequencies. Each point represents a distinct indel pattern. Pearson correlation coefficients indicate concordance between prediction and outcome. Log10-transformed frequencies are shown; a value of 1 was added to all predicted and observed frequencies to avoid log(0). e Comparison of editing efficiencies between mouse blastocysts and mESCs. f Phenotypic comparison of edited (limbless) and wild-type E15.5 embryos. g Editing efficiency of E15.5 embryos. h Mutation patterns of each embryo; blue indicates the MFG predicted using inDelphi. Each column represents a single embryo. i Schematic showing primer positions and the expected long-range PCR amplicon (3991 bp) for the Fgf10 gene. Gel electrophoresis images display PCR products from embryos initially classified as homozygous by short-amplicon NGS, including wild-type and edited embryos, alongside a molecular size marker. Each lane represents an individual embryo. j Pie charts showing the frequencies of major indel genotypes at E15.5 embryos, as predicted by inDelphi, and in mESCs. The dominant indels (4 bp deletion, 1 bp insertion, 13 bp deletion, and 1 bp deletion) are shown, with remaining genotypes grouped as “Others”. Scale bar = 2 mm.

In summary, our pipeline is broadly applicable across different genes in mice and enables the efficient generation of genetically homogeneous and predictable embryos.

Proposed pipeline for optimized genome editing in mouse models

Conventional approaches for generating KO animals typically rely on gRNA design tools that prioritize cutting efficiency and minimize off-target effects, often with limited consideration of the resulting genotypes^22,23 (Fig. 5). These gRNAs are sometimes validated for cleavage activity using in vitro assays, such as the EGxxFP system or a Cas9-mediated cleavage assay. In other cases, they are used directly in downstream experiments. Such strategies often lead to heterogeneous and unpredictable mutant outcomes. Alternatively, our approach employs inDelphi to design gRNAs based on predicted mutational outcomes, complemented by off-target analyses using CRISPOR. Multiple gRNAs are selected with high scores in inDelphi, along with high MIT and CFD specificity scores from CRISPOR. These candidates are first evaluated in vitro using mESCs to assess the editing efficiency and genotypic profiles. An optional EGxxFP assay may also be performed to confirm cutting activity. This screening process enables the selection of the most effective gRNAs for in vivo applications, thereby improving the precision of genotypic outcomes and substantially reducing the number of animals required, in accordance with ethical research principles.

Fig. 5. Proposed pipeline for optimized genome editing in mouse models.

Schematic overview of the regular and proposed workflow for generating genetically modified mice. In the proposed pipeline, candidate gRNAs are first evaluated using inDelphi to predict mutation outcomes and CRISPOR to assess potential off-target effects. Selected gRNAs are then tested in mESCs to evaluate editing efficiency and mutation patterns. An optional EGxxFP assay may be used to further confirm editing efficiency. The gRNA with the optimal balance of efficiency and specificity is subsequently used for animal experiments. This pipeline enables the efficient generation of homogeneous and predictable genome-edited mice in the F0 generation. Created in BioRender. Okamura, E. (2026) https://BioRender.com/l30xep1.

Discussion

We applied inDelphi to mouse embryos targeting the Tyr gene and observed moderate accuracy in predicting mutation patterns at the blastocyst stage, underscoring the need for a more robust predictive system. To address this limitation, we incorporated an in vitro screening step using mESCs, and the mESC-based validation substantially improved the predictive performance of the inDelphi model. By applying this hybrid approach to the Tyr and Fgf10 genes, we successfully generated F0 mutant mice with highly uniform and predictable genotypes. This strategy refines the gRNA selection process and facilitates the direct generation of genotype-specific models in the F0 generation.

inDelphi is a machine learning model trained on large datasets from cultured human and mouse cells²⁵. In this study, we used the mESC-trained version of inDelphi for all gRNA predictions. Nevertheless, we consistently observed that actual genome editing outcomes in mESCs showed stronger concordance with mutation patterns in blastocysts and post-implantation embryos than those predicted by inDelphi alone. The reason for this discrepancy remains unclear, but one possible explanation is that inDelphi may require further refinement or additional training using mESC data generated under experimental conditions that more closely reflect in vivo environments. In the original study²⁵, the training data for the mESC model were derived from editing outcomes at 55-bp human genomic fragments integrated into the mESC genome, which may not replicate endogenous chromatin context or repair process.

Another limitation of inDelphi is its inability to predict editing efficiency. While several computational tools are available for estimating editing efficiency^{35,37,45–47}, our results indicated mESC-based assays offer superior predictive performance, underscoring the value of integrating in vitro data into the gRNA selection process. These findings also highlight the need to further improve existing in silico tools or develop novel ones that can more accurately capture both mutational outcomes and editing efficiencies under biologically relevant conditions.

CRISPR/Cas9 mutational outcomes are shaped not only by target sequence features but also by system-dependent Cas9 exposure kinetics and cellular DNA repair environments (Supplementary Fig. 12). In commonly used reporter assays in transformed somatic cell lines (e.g., HEK293), plasmid-based Cas9 delivery results in sustained Cas9 expression, enabling repeated cycles of cleavage and repair^48,49. In differentiated mammalian cells, DSBs are repaired predominantly by NHEJ, with additional contributions from alternative pathways depending on cellular context^50,51. Consistent with this, large-scale profiling studies have shown that CRISPR-induced indel patterns in somatic cell lines are diverse yet reproducible and reflect contributions from multiple end-joining mechanisms in a sequence-dependent manner⁵². In this study, mESCs were edited by plasmid lipofection, resulting in sustained Cas9 exposure within a pluripotent cellular context. Pluripotent stem cells exhibit DNA damage responses distinct from those of differentiated cells, including differences in checkpoint regulation and repair pathway engagement^19,53,54. More broadly, polymerase θ (POLQ) mediates MMEJ in mammalian cells and contributes to end-joining diversity in a context-dependent manner¹⁹, providing a mechanistic basis for microhomology-associated outcomes. In early embryos, Cas9 is typically delivered as a ribonucleoprotein complex by electroporation, resulting in transient Cas9 activity^48,49. DSB repair in this setting occurs within a developmentally constrained context characterized by rapid cell cycles and developmentally regulated DNA damage responses, and the timing and duration of Cas9 activity strongly influence mutational outcomes, consistent with repair dynamics distinct from those of cultured cells^55–57. Collectively, these differences underscore the importance of Cas9 exposure kinetics and cellular repair context in shaping genome-editing outcomes and provide a framework for comparing mESCs, embryos, and reporter assays in transformed somatic cell lines, while contextualizing our observation that mutational outcomes in mESCs and fertilized embryos are strongly correlated.

Traditional strategies for generating KO animals often prioritize the selection of gRNAs that maximize editing efficiency while minimizing off-target effects^22,23. Although such gRNAs induce DSBs efficiently at target sites, variability in DNA repair outcomes is frequently overlooked, resulting in mosaicism in F0 animals. For example, Sunagawa et al.¹⁵ reported a 100% KO efficiency using a triple-target CRISPR approach against a single gene; however, mosaicism remained prevalent in the F0 generation. Previously, we generated a cynomolgus monkey model of autosomal dominant polycystic kidney disease model using CRISPR. The edited animals harbor diverse insertion/deletion mutations in the PKD1 allele, which resulted in variable phenotypic outcomes⁴. Knock-in (KI)^58,59, base-editing (BE)^60,61, and prime editing (PE)^62,63 technologies enable precise genotype control, but they are generally associated with lower efficiencies and increased technical complexity. Our approach demonstrates that gRNA can be identified that predominantly induce MMEJ-mediated mutations, enabling efficient and homogeneous editing when appropriate microhomology sequences are present near the target site. This strategy offers a valuable complement to conventional KO, KI, BE, and PE approaches for achieving predictable and uniform genetic outcomes. Moreover, it can be integrated with knock-in strategies by enabling pre-testing of key donor design parameters—such as homology arm length, strand orientation, and incorporation of PAM-disrupting or protospacer-altering mutations to prevent re-cutting—as well as by allowing early evaluation of HDR efficiency in mESCs. For instance, Chen et al.⁶⁴ demonstrated a substantial improvement in knock-in efficiency by inhibiting NHEJ, thereby shifting DSB repair toward MMEJ, and further showed that suppressing MMEJ can enhance HDR. Building on these insights, our approach enables pre-validation of MMEJ-dominant gRNAs, allowing researchers to identify gRNAs that intrinsically favor MMEJ-mediated repair prior to embryo manipulation. This strategy reduces the time and labor required for gRNA screening and supports the more efficient design of MMEJ-based genome editing experiments.

It is important to note that even frameshift mutations do not necessarily replicate the exact sequence of human pathogenic variants. This variability limits the ability of animal models to accurately mimic human diseases. Notably, among all human pathogenic variants, 58% are point mutations and 25% are deletions⁶⁵. Of these deletions, 57% are thought to arise through MMEJ⁶⁶. Given its prevalence, MMEJ offers a unique opportunity to generate animal models with precise, disease-relevant genotypes that closely reflect human pathogenic deletion-type mutations.

Finally, the 3Rs—Replacement, Reduction, and Refinement—are fundamental ethical principles in animal research, emphasizing the minimization of animal harm. However, generating CRISPR KO mice remains a stochastic process owing to the unpredictability of DNA repair outcomes. This inefficiency often necessitates the sacrifice of numerous F0 animals because the desired genotype typically emerges only in the F1 generation. To address this issue, our approach uses inDelphi to design gRNAs based on predicted mutational outcomes. By first testing these gRNAs in cell culture, we can evaluate the editing efficiency and mutation patterns to select the most effective candidates for in vivo application. This strategy improves the precision of genotypic outcomes while substantially reducing the number of animals required, thereby aligning with ethical research principles.

In larger animal models, such as pigs and non-human primates, extended gestation periods and delayed sexual maturity further complicate the generation of genetically uniform animals^67,68. Establishing species-specific ESC or iPSC systems may provide an equivalent in vitro testing platform to refine editing strategies prior to in vivo experiments. In these species, achieving the desired genotype directly in the F0 generation is particularly critical for reducing time, cost, and animal use. By facilitating predictable genotype generation in the F0 generation, our approach has the potential to accelerate translational research using large animal models, especially in the context of disease modeling and preclinical testing. Continued advances in genome editing technologies are therefore essential to enhance the precision and efficiency of F0 editing in these challenging models.

Materials and methods

Analysis of inDelphi-based indel prediction and editing efficiency prediction tools

For the inDelphi (https://indelphi.giffordlab.mit.edu/batch) analysis, the target exon sequences were used as the sequence context and submitted using the batch input function²⁵. For the Cas9 PAM and Cell type setting, NGG and mouse embryonic stem cell (mESC) were selected, respectively. The analysis yielded predictions of indel frequency distributions, including calculated precision scores, microhomology strength, and frameshift frequencies for all possible gRNA candidates within the input sequence. For the evaluation of gRNA editing efficiency, CRISPRon (https://rth.dk/resources/crispr/crispron/), CRISPick (https://portals.broadinstitute.org/gppx/crispick/public), and CRISPRscan (https://www.crisprscan.org) were used. These tools predict gRNA efficiency based on computational models that incorporate sequence-derived features, such as nucleotide composition and position-specific scoring. Guide sequences were submitted to each platform, and efficiency scores were obtained to assess and prioritize gRNA candidates based on predicted editing activity.

EGxxFP assay

The EGxxFP assay was performed using HEK293FT cells to assess genome editing efficiency as described previously³⁷. A reporter plasmid was constructed by inserting a part of the mouse Tyr exon 1 sequence into a disrupted EGFP coding sequence of the pCAG-EGxxFP plasmid. Oligonucleotide sequences used for this insertion are listed in Supplementary Table 2. The reporter plasmid was co-transfected with the PX459 plasmid⁶⁹ containing Cas9 and a puromycin-resistant gene as well as gRNA expression cassettes using Lipofectamine 2000 (Thermo Fisher Scientific, Waltham, MA, USA). gRNAs targeting Tyr exon 1 were cloned into PX459 using annealed oligonucleotides, as listed in Supplementary Data. After 48 hours, the GFP fluorescence signal intensity was assessed using a fluorescence microscope (BZ-9000; Keyence, Osaka, Japan) and BD FACSCalibur flow cytometer (BD Biosciences, San Jose, CA, USA) (Supplementary Fig. 1). pCAG-EGxxFP was a gift from Masahito Ikawa (Addgene plasmid #50716; http://n2t.net/addgene:50716; RRID:Addgene_50716). pSpCas9(BB)-2A-Puro (PX459) V2.0 was a gift from Feng Zhang (Addgene plasmid #62988; http://n2t.net/addgene:62988; RRID:Addgene_62988).

Genome editing in mESCs

C57BL/6 J mESCs⁷⁰ were obtained from Riken BioResource Research Center (Ibaraki, Japan). mESCs were cultured as described previously⁷¹. In brief, ESC medium [DMEM] (Thermo Fisher Scientific, 10565-018) supplemented with KSR, 1 mM sodium pyruvate, 0.1 mM 2-mercaptoethanol (Sigma-Aldrich, St. Louis, MO, USA, M3148), 1× non-essential amino acids (Thermo Fisher Scientific, 11140-050), 1 mM l-glutamine, 1000 U leukemia inhibitory factor/ml, 3 mM CHIR99021, and 1 mM PD0325901 was used.

For transfection, PX458 plasmids⁶⁹ harboring the gRNA and Cas9-2A-EGFP expression cassettes were introduced for 12 × 10⁵ mESCs in a 24-well plate using Lipofectamine 2000 (Thermo Fisher Scientific). Oligonucleotides used for gRNA cloning into PX458 are listed in Supplementary Data. Cells were cultured for 48 hours and then incubated with 7-AAD to assess cell viability. Fluorescence-activated cell sorting (FACS) was performed using the FACSAria Fusion flow cytometer (BD Biosciences) to isolate GFP-positive and 7-AAD-negative cells. Genomic DNA was subsequently extracted using the NaOH lysis method. pSpCas9(BB)-2A-GFP (PX458) was a gift from Feng Zhang (Addgene plasmid #48138; http://n2t.net/addgene:48138; RRID:Addgene_48138).

Animal experiments

All experimental procedures were approved by the Animal Care and Use Committee of Shiga University of Medical Science (approval number: 2024-5-20(H1)). C57BL/6JJmsSlc and Slc:ICR mice were purchased from Japan SLC (Shizuoka, Japan). The mice were kept in plastic cages under specific pathogen-free conditions, with the room temperature maintained at 23.0 ± 3.0 °C, 50.0 ± 20.0% relative humidity, and a 12:12 hour light/dark cycle. Fertilized eggs were collected through in vitro fertilization⁷². In brief, oocytes were retrieved from the oviducts of 3-week-old C57BL/6 JJmsSlc females that had undergone superovulation via intraperitoneal administration of CARD HyperOva (Kyudo, Tosu, Japan), followed by human chorionic gonadotropin treatment. Sperm were obtained from the caudal epididymis of 8- to 16-week-old C57BL/6JJmsSlc males and preincubated in Fertiup Mouse Sperm Preincubation Medium (Kyudo). Fertilization was performed in HTF medium (Ark Resource, Kumamoto, Japan), followed by incubation at 37 °C in a 5% CO₂ atmosphere for 3 to 6 hours. The fertilized eggs were then washed to remove cumulus cells and sperm and subsequently cultured in KSOM medium (Ark Resource) until electroporation.

Electroporation of fertilized eggs was performed in Opti-MEM I medium (Thermo Fisher Scientific) containing the RNP complex (crRNA:tracrRNA duplex and HiFi-Cas9 protein; Integrated DNA Technologies, Coralville, IA, USA) using a Super Electroporator NEPA21 (Nepa Gene, Ichikawa, Japan) with a 1 mm gap platinum electrode (BEX, Tokyo, Japan). The poring pulse was set as follows: 30–40 V, 3.5 ms pulse width, 50 ms pulse interval, 4 pulses, 10% attenuation rate, and + polarity. The transfer pulse was set as follows: 5 V, 50 ms pulse width, 50 ms pulse interval, 5 pulses, 40% attenuation rate, and ± polarity. The electroporated one-cell embryos were cultured in KSOM medium until the two-cell stage and then transferred to pseudopregnant ICR females or cultured to the blastocyst stage (E4.5). Embryonic day E11.5 or E15.5 embryos were dissected from the uteri of euthanized surrogate mothers, and genomic DNA was extracted using the phenol-chloroform method, with retinal pigmentation quantified by background-subtracted mean grayscale intensity. For a blastocyst embryo analysis, embryos were collected in a single tube, with an average of 30 embryos per sample. Genomic DNA was extracted using the NaOH lysis method.

No measures to minimise confounding factors were used because embryos were grouped only by gRNA and handled under identical experimental conditions. Blinding was not applied as phenotypes were objective and sequencing analysis was automated. No a priori power calculation was performed; sample sizes were determined based on expected rates of IVF success, embryo development, and embryo-transfer. No humane endpoints were defined because all animals were euthanised at predetermined experimental timepoints and no survival procedures were conducted.

Target amplification by PCR, library preparation, and iSeq 100 sequencing

Target amplification for amplicon-seq was performed with nested PCR using a T100 thermal cycler (Bio-Rad, Hercules, CA, USA). Primer sets were designed to amplify a ~ 300 bp region flanking the CRISPR cutting site for each target, as listed in Supplementary Data. In the second PCR step, dual-indexing primers were used, custom-synthesized by Eurofins Genomics (Tokyo, Japan) based on publicly available Illumina adapter sequences. PCR products were analyzed via 2% agarose gel electrophoresis to confirm amplification and subsequently purified using a Gel/PCR Purification Mini Kit (Favorgen Biotech Corp., Ping-Tung, Taiwan) and AMPure XP bead purification system (Beckman Coulter Life Sciences, Indianapolis, IN, USA). The purified libraries were loaded onto the iSeq 100 system (Illumina, San Diego, CA, USA) for next-generation sequencing. Sequencing results were analyzed using CRISPResso2³⁶ (https://crispresso2.pinellolab.org/submission) to assess genome editing outcomes.

Multiplex genotyping with nanopore sequences

The purified genomic DNA was amplified by PCR using KOD Multi & Epi (Toyobo, Osaka, Japan) and target amplicon primers with the universal sequence (22-mer) located on the 5′ side (Supplementary Data, 1st primer). Fivefold dilutions of the PCR products were used as templates for nested PCR performed using KOD Multi & Epi and barcode attachment primers (Supplementary Data: 2nd primers (barcode)). Equal amounts of barcoded PCR products were mixed and column-purified using FastGene Gel/PCR Extraction Kit (Nippon Genetics, Germany). The concentration of PCR product was adjusted to 15 ng/μL. The library was prepared using Ligation Sequencing kit (SQK-LSK114; ONT, Oxford, UK), NEBNext® Companion Module for Oxford Nanopore Technologies® Ligation Sequencing (NEB, E7180S) and NEBNext® Quick Ligation Module (NEB, E6056S) in accordance with the manufacturer’s instructions. The prepared library was loaded into the Flow Cell (R10.4.1) and run with PromethION for 72 h. Base-calling and barcode demultiplexing were performed by dorado (version 1.3.0).

Genetic analysis using DAJIN2

DAJIN2 (version 0.7.4) analyzes demultiplexed BAM files to identify alleles within long-range on-target regions⁴³. It classifies nanopore sequencing reads according to their mutation profiles. First, DAJIN2 aligns sequences from the user-provided FASTA file using minimap2 via mappy (version 2.26)⁷³. The aligned sequences are then converted into the MIDSV format (Match, Insertion, Deletion, Substitution, and inVersion) to extract base-level mutation profiles (https://github.com/akikuno/midsv). Next, the mutation rate of the control is subtracted from that of the sample, and the resulting difference matrix is clustered. The resulting cluster labels are assigned as allele identifiers. Subsequently, pysam (version 0.21)⁷⁴ generates sorted BAM files for each allele. The corresponding genomic coordinates and chromosome lengths, based on the input FASTA file and genome assembly ID, are retrieved from the UCSC Genome Browser. DAJIN2 then replaces the chromosome name (SN) and length (LN) fields in the BAM headers accordingly. Finally, consensus sequences for each allele are generated as FASTA and HTML files using cstag (version 1.0.5)⁷⁵. Allele groups representing less than 5% of the total are merged and annotated as “Others”.

Statistical analysis

To compare inDelphi predictions with mouse experimental data, wild-type reads were excluded from the sequencing results. The remaining reads were then normalized to 100% to determine the observed frequencies of each retained variant. Correlations were assessed between mutation patterns observed in blastocysts and inDelphi predictions as well as between blastocyst data and mESC results. A linear mixed-effects model was applied, and Welch’s t tests using Satterthwaite’s method for denominator degrees of freedom was conducted. Hotelling’s T² test was used to assess differences between inDelphi predictions and mESC experimental results. The T² statistic was converted to an F-statistic, and the corresponding p-value was calculated. A value of p < 0.05 was considered statistically significant. Microhomology lengths were grouped into three categories (1 bp, 2–3 bp, ≥4 bp). Deletion-size distributions were compared using the Kruskal–Wallis test with pairwise Wilcoxon rank-sum tests (Benjamini–Hochberg correction). Violin and box plots were generated in R (ggplot2), with significance annotated using ggpubr. Statistical analyses were conducted using Python (NumPy, SciPy), R, and GraphPad Prism software.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

E.O. K.W. and M.E. designed the study. E.O. and K.L. designed the gRNAs, performed cell experiments, and conducted mouse experiments. T. T., H. S., A. K., Y. I., and S. M. performed the sequencing analysis. E. O. K.L. K. W., S. M., Y. I., and K. W. analyzed the data. K.L., E.O., S. M. K. W., and M.E. co-wrote the paper. All authors edited the manuscript. E.O., S. M. K. W., and M.E. supervised the project.

Peer review

Peer review information

Communications Biology thanks Tomoji Mashimo and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Mo Li and Mengtan Xing. A peer review file is available.

Data availability

All data supporting the findings of this study are available within the article and its supplementary information files.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s42003-026-09771-z.