Coiled-coil heterodimer-mediated split base editing systems enable flexible and robust nucleotide substitutions

Abstract

Base editors (BEs) enable precise base substitutions, but their size exceeds the packaging capacity of adeno-associated virus (AAV), impeding in vivo applications. Here we design a split BE system that recruits deaminases to Cas9 nickase via coiled-coil heterodimers, resulting in various coiled-coil heterodimers-mediated base editors (CC-BEs), including cytidine base editor (CC-CBE), adenine base editor (CC-ABE), and their derivatives. We reveal that CC-BEs maintain and even improve the editing efficiency of the original unsplit BEs across various cell types and editing scopes, achieving maximum enhancements of 9.6-fold in human immortalized cells and 12.4-fold in primary somatic cells for CC-CBE. Using CC-ABE, we validate in vivo editing efficiency and successfully achieve A-to-G conversion in the Pcsk9 and Dmd genes via dual-AAV vectors in mice. Altogether, we develop a simple and universal strategy to address the challenges posed by the large size of BEs without compromising editing efficiency for base substitutions in vivo.

Base editors (BEs) enable precise base substitutions but are limited by their large size. Here, the authors engineer a split BE system utilizing coiled-coil heterodimers (CC-BE) and demonstrate that CC-BEs maintain or even enhance efficiency, enabling gene therapy applications via dual AAV.

Introduction

As a precise gene editing tool, the DNA base editor is constructed based on the CRISPR/Cas system in conjunction with deaminases, glycosylases, or their derivatives, enabling base pair substitutions without inducing double-strand DNA breaks or necessitating exogenous donor templates^1,2. Currently, a variety of base editors have been developed to enable multiple types of base pair substitutions, such as ABE (A-to-G)³, CBE (C-to-T)^4,5, CGBE (C-to-G)^6–8, AYBE (A-to-Y)⁹, DAF-BEs (T-to-G, C-to-G)¹⁰, and TSBE (T-to-G, T-to-C)^8,11. Base editing technology demonstrates significant potential applications in crop breeding¹², generation of animal models relevant to human diseases^13–15, genetic screening¹⁶, and lineage tracing^17,18, particularly offering substantial possibilities for the treatment of numerous genetic diseases associated with pathogenic point mutations¹⁹.

Given its relatively low pathogenicity, low immunogenicity, broad tissue tropism, mature production processes, and ability to establish long-term gene expression, adeno-associated virus (AAV) emerges as a potent platform for gene therapy delivery^20,21 compared to other delivery tools like lipid nanoparticles (LNPs)²², virus-like particles (VLPs)²³, enveloped delivery vehicles (EDVs)²⁴, and selective endogenous encapsidation for cellular delivery (SEND)²⁵. However, the AAV can only package cargo up to about 4.7 kb, inherently smaller than the size of SpCas9-based base editors^26,27, impeding their therapeutic application. To bypass this limitation, intein-mediated split base editors with dual-AAV strategy^27–31 and miniature base editors^32–36 delivered via a single AAV particle have been developed. These strategies have their own limitations in terms of application. Split inteins self-assemble and undergo the ligation of the N- and C-termini of a functional protein through post-translational protein trans-splicing, thereby forming a fully functional protein and excising the intein in the process³⁷. However, the split intein-mediated protein trans-splicing strategy could potentially alter protein function due to the adjacent residues at the extein/intein boundaries on which inteins heavily rely, which are left behind as “footprints”³⁸. Another disadvantage is that selecting the split sites of functional proteins is challenging and may compromise the efficiency of the protein^39,40. Miniature nucleases, such as Cas12f, TnpB, and IscB, are utilized to develop miniature base editors through combinations with cytosine and adenine deaminases^{32–35,41,42}. Nevertheless, these miniature gene editing tools exhibit low editing activity and limited editing scope, fundamentally restricting precision targeting^43,44. Therefore, there is an unmet need to develop a flexible and efficient base editing system for in vivo single-base substitutions.

Heterodimeric coiled-coil (CC) peptides are typically composed of two alpha helices winding around each other through hydrophobic and electrostatic interactions, enabling highly specific and tight dimerization^45,46. Dimerization of the base editor system via CC peptides can offer distinct advantages over the existing split strategies, as this approach does not require selecting splitting sites for the nuclease or impose an additional burden to the AAV vector capacity. Here, we develop a CC-BE system that recruits deaminase to Cas9 nickase via CC dimer-forming modules and demonstrate that CC-BE can serve as a general base substitution platform for conferring high efficiency and flexibility. We generate 5 different versions of CC-BEs, including CC-ABE8e, CC-BE3, CC-ABE9, CC-TadCBE, and CC-AYBE. All 5 versions of CC-BEs exhibit similar or even higher efficiencies of base editing compared to the unsplit BE at most of the loci we tested. Furthermore, CC-BEs enhance or maintain the editing efficacy of unsplit BE across various cell types (HEK293T, MEF, and PFF) and targeting scopes (NGG PAM and NG PAM), robustly demonstrating the versatility of the CC-BE system. Finally, we evaluate the in vivo application of CC-BE using dual-AAV vectors to edit the mouse Pcsk9 gene and correct the Dmd gene, demonstrating high targeted base editing efficiencies of up to 79.0% and 21.9%, respectively. This validates their considerable potential as a simple, flexible, and efficient gene-editing tool for therapeutic applications.

Results

Development of CC-CBE with high editing activity by coiled-coil heterodimers

To advance towards a split cytosine base editor, we split CBE into two parts (APOBEC3 and nCas9-UGI) and each part was fused to coiled-coil dimer-forming peptides, P3–P4 pair or N5–N6 pair, so that the N- and C-terminus of CBE were subject to AAV packaging size restriction. We fused P3 or N5 to the C-terminus of APOBEC3 and P4 or N6 to the N-terminus of nCas9 (Cas9 nickase, D10A) (Fig. 1a, b). AlphaFold 3 3D structure predictions⁴⁷ showed that in the presence of two sections, the pair P3 and P4 or N5 and N6 heterodimerizes, bringing APOBEC3 and nCas9-UGI into proximity and facilitating their assembly into a functional unit (Fig. 1a). As the control, a direct split CBE system without affinity modules and the previously mentioned split-intein CBE system²⁷, in which the intein splices at Cys 574 or Ser 714, were constructed (Fig. 1b).

Fig. 1. CC-CBE permits efficient precise base editing across genomic sites in HEK293T cells.

a 3D structure comparison of CBE and CC-CBE based on AlphaFold 3 prediction. b Schematics of CBE, sCBE-P3P4, sCBE-N5N6, and sCBE-ctrl. c, d Comparison of target C-to-T precise editing (c) and undesired indel efficiencies (d) between CC-CBEs, direct split CBE, split-intein CBEs, and unsplit CBE system. Data represent mean ± standard deviation of 4 independent biological replicates. e–g Precise base editing (e, f) and undesired indel efficiencies (g) of sCBE-P3P4, sCBE-N5N6, sCBE-ctrl, and unsplit CBE system across 9 tested loci. Data represent mean ± standard deviation of 4 independent biological replicates. h Parallel comparison of average base editing activities by sCBE-P3P4, sCBE-N5N6, sCBE-ctrl, and unsplit CBE system. Statistical significance was determined via two-tailed Student’s t tests. Exact P values are provided in Source data file. (*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, ns indicating not significant). Source data are provided as a Source data file.

To more accurately assess editing efficiency and eliminate the impact of transfection efficiency, we introduced EGFP into the N-terminal vector of CC-BE, mCherry into the C-terminal vector of CC-BE, and EGFP into the unsplit BE vector. Fluorescence-activated cell sorting was performed to evaluate base editing efficiency after transfection (Supplementary Fig. 1a). Untreated cells were used as a control for setting the gates for cell sorting (Supplementary Fig. 1b). For the sBE-P3P4, sBE-N5N6, and sBE-ctrl groups, mCherry and EGFP double-positive cells were collected; for the unsplit BE group, EGFP-positive cells were collected for subsequent analysis (Supplementary Fig. 1c). We first compared editing efficiency of CC-CBE with unsplit CBE, direct split CBE, and split-intein CBE across 2 endogenous loci in sorted HEK293T cells (Fig. 1c). The results showed that the precise editing efficiencies of sCBE-P3P4 and sCBE-N5N6 were higher than those of unsplit CBE and direct split CBE (Fig. 1c). And compared to split-intein CBE systems, CC-CBE showed superior or similar efficiencies (Fig. 1c). As we further investigated, the levels of undesirable byproducts of CC-CBE were similar or slightly lower than those of intact CBE (Fig. 1d). Accordingly, these data suggest that the performance of CC-CBE is superior to or comparable with the widely used split CBE strategies. To demonstrate that two CBE components can reassemble into an active form upon coiled-coil heterodimers, the direct split CBE system without affinity modules was used as a control (hereafter sCBE-ctrl) in subsequent studies.

We next examined the universality of CC-CBE at 9 endogenous loci to proceed to a broader scale (Fig. 1e–g). HEK293T cells were transfected with the unsplit CBE, CC-CBE, or directly split CBE system with various sgRNAs to target 9 sites, then sorted to evaluate base editing efficiencies. Consistent with the results presented above, CC-CBE exhibited significantly higher on-target activity at all tested loci compared to the unsplit CBE and sCBE-ctrl (Fig. 1e, f). Among the 9 genomic sites tested, the sCBE-P3P4 demonstrated a mean C-to-T conversion rate of 69.8% (ranging from 41.9% to 90.6%), achieved an average 2.8-fold improvement compared to intact BE3 (ranging from 6.3% to 38.2%, averaging 24.9%) (Fig. 1e). The maximum efficiency improvement for sCBE-P3P4 was 9.6-fold at RNF2 site 2 (from 6.3% to 60.4%) (Fig. 1e). The sCBE-N5N6 exhibited a mean C-to-T conversion rate of 71.4% (ranging from 44.5% to 92.7%), achieved an average 2.9-fold improvement relative to intact BE3 with a maximum improvement of 7.1-fold at RNF2 site 2 (from 6.3% to 44.5%) (Fig. 1e). And the indel frequencies of CC-CBE were similar or lower than those of intact CBE at 4 loci (Fig. 1g). In contrast, the direct split CBE showed unstable performance in terms of both editing efficiencies and indels (Fig. 1e–g). We next examined the editing windows of CC-CBE, direct split CBE, and intact CBE. The observed editing window for CC-CBE across all loci spanned position 3–13, while direct split CBE displayed unpredictable pattern owing to the absence of affinity modules. (Fig. 1f, h). These results showed that CC-CBE exhibited broader editing window compared to unsplit CBE, thereby extending the utility of CBE beyond the original editing window (position 4–8).

To provide a more comprehensive characterization of CC-CBE, we analyzed its base editing performance in transfected bulk HEK293T cells (Supplementary Fig. 2a–d). The data showed that sCBE-P3P4 exhibited higher precise editing efficiencies (range from 12.65% to 28.11%) than original unsplit CBE (range from 6.08% to 17.37%) at five out of six loci, while sCBE-N5N6 achieved higher efficiencies (range from 9.83% to 33.70%) at all the tested loci (Supplementary Fig. 2a, b). The sCBE-P3P4 induced lower indel frequencies than the unsplit CBE at all six loci, while sCBE-N5N6 showed variable indel frequencies across loci yet remained within a tolerable range (Supplementary Fig. 2c). A broad editing window of CC-CBE was observed in bulk transfected cells (Supplementary Fig. 2d). The data presented here were largely consistent with those obtained from sorted cells.

We noticed that the constructs for CC-BE contain a total of 4×NLS, whereas the unsplit BE contains only 2×NLS. We therefore hypothesized that the differences in editing efficiency between CC-CBE and CBE were due to the varying numbers of NLS. To test this hypothesis, we constructed an unsplit CBE containing 4×NLS for the evaluation of editing efficiency (Supplementary Fig. 3a). Our results indicated that the C-to-T efficiency of 4×NLS-CBE was slightly higher than or comparable to that of 2×NLS-CBE at 3 loci out of 4 edits; however, it still did not exceed that of CC-CBE (Supplementary Fig. 3b). The indel frequencies of 4×NLS-CBE were similar to or higher than those of 2×NLS-CBE (Supplementary Fig. 3c). This suggests that the increased editing efficiency of CC-CBE is not attributable to differences in the number of NLS. We also compared the Cas9 protein expression levels between CC-CBE and intact CBE using Western blot analysis (Supplementary Fig. 4). No obvious differences were observed, suggesting that the differences in editing efficiency may not arise from variations in Cas9 protein expression levels. AI-based prediction of the 3D protein structures of intact CBE and CC-CBE could be used to better elucidate the variances in editing efficiencies (Fig. 1a). Due to the presence of coiled-coil heterodimers, the base editor was reconstituted, bringing UGI and APOBEC3 together compared to intact CBE. We reasoned that the conformational change contributed to a particularly noticeable change in the overall performance of the CC-CBE. Therefore, the results above support the conclusion that CC-CBE demonstrates a highly adaptable strategy for enhancing the efficiency of CBE.

CC-ABE system exhibits high editing efficiency

To facilitate the flexibility of adenine base editor, we constructed CC-ABE system by splitting ABE into two fragments using coiled-coil heterodimers. As ABE8e is the most efficient and versatile ABE variant⁴⁸, we first chose this to test our CC-ABE system. Through the structure predicted by AlphaFold 3, we found that coiled-coil heterodimers allow TadA8e and nCas9 to dimerize at the target site (Fig. 2a). In the design of CC-ABE8e, the C-terminus of TadA8e was fused to P3 or N5 and P4 or N6 was attached to the N-terminus of nCas9 (Fig. 2b). We designated these ABE systems as sABE8e-P3P4 and sABE8e-N5N6, respectively (Fig. 2b). As a control for CC-ABE8e system, we devised the direct split ABE (sABE8e-ctrl) omitted the respective affinity modules (Fig. 2b). To investigate the performance of CC-ABE in sorted HEK293T cells, we compared its editing efficiency with that of intact ABE8e, sABE8e-ctrl, and split-intein ABE8e (Npu C574, Npu S714) at three human endogenous genomic sites (Fig. 2c). We observed that CC-ABE8e achieved at least equivalent level of precise editing efficiencies compared to Npu C574, Npu S714, intact ABE8e and sABE8e-ctrl at two loci. sABE8e-P3P4 exhibited better performance compared to Npu S714 and sABE8e-ctrl, while showing lower efficiency to Npu C574 and intact ABE8e at FANCF site 3 (Fig. 2c). The undesired indel byproducts generated by CC-ABE8e were lower than those of sABE8e-ctrl, intact ABE8e, and split-intein ABE8e at 2 of the 3 target sites (Fig. 2d).

Fig. 2. CC-ABE enables efficient base editing at multiple endogenous loci in HEK293T cells.

a Predicted structure of unsplit ABE, CC-ABE, and their alignment by AlphaFold 3. b Schematics of ABE, CC-ABE, and direct split ABE. c, d Comparison of the efficiencies of A-to-G conversion (c) and indel rates (d) generated by sABE-P3P4, sABE-N5N6, sABE-ctrl, intact ABE, and split-intein ABE (Npu C574, Npu S714) at 3 loci. Data represent mean ± standard deviation of 4 independent biological replicates. e–g The A-to-G editing efficiency (e, f) and undesired indel efficiencies (g) of sABE-P3P4, sABE-N5N6, sABE-ctrl, and intact ABE were examined at 13 loci. Data represent mean ± standard deviation of 4 independent biological replicates. h Average A-to-G editing efficiency of sABE-P3P4, sABE-N5N6, sABE-ctrl, and intact ABE at 13 target sites in f. Statistical significance was determined via two-tailed Student’s t tests. Exact P values are provided in Source data file. (*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, ns indicating not significant). Source data are provided as a Source data file.

Next, we compared the editing efficiencies of CC-ABE8e with sABE8e-ctrl and intact ABE8e at 13 human genomic loci in sorted HEK293T cells, aiming to explore the performance of CC-ABE8e across a wider range (Fig. 2e–h). We found that compared to sABE8e-ctrl (3.23% to 95.52%, averaging 64.88%), CC-ABE8e (sABE8e-P3P4, 6.61% to 96.75%, averaging 76.07%; and sABE8e-N5N6, 2.64% to 92.48%, averaging 67.23%) had the higher editing efficiency, which displayed similar level to intact ABE8e (9.80% to 95.12%, averaging 76.13%) (Fig. 2e, f). The undesired indel rates of CC-ABE8e were lower than those of sABE8e-ctrl at 7 of the 13 target sites (Fig. 2g). Compared to unsplit ABE8e, CC-ABE8e showed lower undesired indel rates at 4 loci, and comparable rates at 6 loci (Fig. 2g). Subsequently, the average A-to-G conversion rates within the protospacers of all loci were calculated. The results demonstrated that CC-ABE8e (sABE8e-P3P4 and sABE8e-N5N6) induced a higher frequency of A-to-G conversions at positions A₉–A₁₄ compared to intact ABE and sABE8e-ctrl, while the efficiency at A₅–A₈ remained similar to intact ABE (Fig. 2h).

We next evaluated the base editing of CC-ABE in transfected bulk HEK293T cells (Supplementary Fig. 2e–h). CC-ABE showed higher or minimally comparable editing efficiencies relative to sABE-ctrl, and comparable or lower efficiencies compared with the unsplit ABE (Supplementary Fig. 2e). Meanwhile, sABE-P3P4 demonstrated better performance than sABE-N5N6. (Supplementary Fig. 2e, f). In terms of indels, sABE-P3P4 exhibited lower efficiencies than unsplit ABE at 4 out of 6 loci, while sABE-N5N6 achieved lower efficiencies at 5 out of 6 loci (Supplementary Fig. 2g). The relative A-to-G conversion efficiencies within the protospacers were consistent with those observed in sorted cells (Supplementary Fig. 2h).

The A-to-G efficiency of 4×NLS-ABE was slightly higher than or comparable to that of 2×NLS-ABE, while it was slightly lower than or comparable to that of CC-ABE (Supplementary Fig. 3d). The indel frequencies of 4×NLS-ABE were similar to or lower than those of 2×NLS-ABE (Supplementary Fig. 3e). The Cas9 expression levels of CC-ABE and intact ABE were comparable (Supplementary Fig. 4). Given the high efficiencies observed in previous results, we hypothesize that splitting CC-ABE into two separate components (TadA8e-P3/N5 and P4/N6-nCas9) may increase the off-target efficiency, as both components would exhibit greater flexibility within the genome. We then developed an iterative version of CC-ABE (split Cas9-based CC-ABE), in which the nCas9 was split into two parts before Cys 574, which may potentially decrease Cas9-dependent off-target activities and restrict the deaminase’s flexibility. We compared the editing efficiencies and undesired indels of split Cas9-based CC-ABE8e with those of sABE8e-ctrl and intact ABE8e at the same human genomic loci in bulk and sorted HEK293T cells (Supplementary Fig. 5a–d). The results were consistent with the data shown in Fig. 2. The split Cas9-based CC-ABE displayed generally similar editing efficiencies to the unsplit ABE system without increasing the indels in either bulk or sorted cells (Supplementary Fig. 5a–d). These findings suggest that the editing efficiency of CC-ABE8e is comparable to intact ABE, characterized by low indel rates and higher peak activity within the editing window.

Engineering base editor variants of TadCBE, ABE9, and AYBE with coiled-coil heterodimers

To evaluate the potential of applying our CC-BE system to other base editing tools, we engineered CC-based TadCBE⁴⁹, ABE9⁵⁰, and AYBE⁹. Evolved from TadA8e, TadA-derived cytosine base editors (TadCBEs) exhibit enhanced on-target editing activity, reduced size, and minimized off-target effects⁴⁹. We split the TadCBE into two parts before Cys 574 in the nCas9 domain, then fused each part to coiled-coil heterodimers, resulting in CC-TadCBE (Fig. 3a and Supplementary Fig. 6a). The C-to-T editing efficiency of CC-TadCBE (sTadCBE-P3P4, range 33.40%–80.54%, mean 57.17%; and sTadCBE-N5N6, range 49.51%–82.73%, mean 64.92%) was substantially higher than sTadCBE-ctrl (range 7.67%–20.48%, mean 14.67%) and the unsplit TadCBE (range 9.96%–32.47%, mean 19.75%) (Fig. 3b). Compared to intact TadCBE, the maximum performance improvements achieved were 4.12-fold by sTadCBE-P3P4 at FANCF site 2 and 4.97-fold by sTadCBE-N5N6 at RNF2 site 2 (Fig. 3b and Supplementary Fig. 6d). CC-TadCBE, sTadCBE-ctrl, and intact TadCBE generally showed low residual A-to-G editing across adenosines (Fig. 3c, f). The purity of C-to-T was much higher than that of A-to-G at all target sites (Fig. 3d). CC-TadCBE exhibited a significantly high peak editing efficiency at positions C₃–C₈ compared to intact TadCBE and sCBE-ctrl (Fig. 3e).

Fig. 3. Development of CC-mediated base editors: CC-TadCBE, CC-ABE9, and CC-AYBE.

a Schematics of CC-TadCBE. b, c The C-to-T (b) and A-to-G (c) editing efficiencies of sTadCBE-P3P4, sTadCBE-N5N6, sTadCBE-ctrl, and TadCBE were examined at 4 loci in HEK293T cells. Data represent mean ± standard deviation of 4 independent biological replicates. d The percentage of C-to-T and A-to-G editing of sTadCBE-P3P4, sTadCBE-N5N6, sTadCBE-ctrl, and TadCBE at the same targets as shown in b and c. e, f Average C-to-T (e) and A-to-G (f) editing efficiency of sTadCBE-P3P4, sTadCBE-N5N6, sTadCBE-ctrl, and TadCBE at 4 loci in b and c. g Schematics of CC-ABE9. h The A-to-G editing efficiency of sABE9-P3P4, sABE9-N5N6, sABE9-ctrl, and ABE9 was examined at 4 loci in HEK293T cells. Data represent mean ± standard deviation of 4 independent biological replicates. i The undesired indel efficiencies of sABE9-P3P4, sABE9-N5N6, sABE9-ctrl, and ABE9. Data represent mean ± standard deviation of 4 independent biological replicates. j Average A-to-G editing efficiency of sABE9-P3P4, sABE9-N5N6, sABE9-ctrl, and ABE9 at 4 loci in h. k Schematics of CC-AYBE. l–n The A-to-C (l), A-to-T (m), and A-to-G (n) editing efficiencies of sAYBE-P3P4, sAYBE-N5N6, sAYBE-ctrl, and AYBE were examined at 3 loci in HEK293T cells. Data represent mean ± standard deviation of 4 independent biological replicates. o The undesired indel efficiencies of sAYBE-P3P4, sAYBE-N5N6, sAYBE-ctrl, and AYBE. Data represent mean ± standard deviation of 4 independent biological replicates. p–r Average A-to-C (p), A-to-T (q), and A-to-G (r) editing efficiency of sAYBE-P3P4, sAYBE-N5N6, sAYBE-ctrl, and AYBE at 3 loci in l–n. s The percentage of A-to-G, A-to-C, and A-to-T editing of sAYBE-P3P4, sAYBE-N5N6, sAYBE-ctrl, and AYBE at the same targets as shown in l–n. Statistical significance was determined via two-tailed Student’s t tests. Exact P values are provided in Source data file. (*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, ns indicating not significant). Source data are provided as a Source data file.

ABE9 precisely edits adenines within a 1–2 nucleotide window and demonstrates the highest efficiency at A₅⁵⁰. We fused coiled-coil heterodimers to the C-terminal of TadA9 and N-terminal of nCas9, resulting in CC-ABE9 (Fig. 3g and Supplementary Fig. 6b). The A-to-G editing efficiency of CC-ABE9 was improved when compared with sABE-ctrl at 4 target sites, and was comparable to or slightly lower than intact ABE9 at 3 of the 4 loci (Fig. 3h and Supplementary Fig. 6e). The undesired indel byproducts generated by CC-ABE9 were lower than or similar to those of sABE-ctrl and intact ABE9 at 2 of the 4 loci (Fig. 3i). As shown in Fig. 3J, CC-ABE9 precisely edited the A₅ position within the protospacer sequence, which was consistent with the previous study⁵⁰.

AYBE can achieve A-to-C and A-to-T transversion editing through the fusion of ABE with human N-methylpurine DNA glycosylase protein (MPG)⁹. We divided the AYBE into two fragments before Cys 574 in the nCas9 domain, fused each part to coiled-coil heterodimers, and generated CC-AYBE (Fig. 3k and Supplementary Fig. 6c). The editing efficiency and A-to-Y purity of CC-AYBE were higher than sAYBE-ctrl and comparable to, or lower than intact AYBE at all target sites (Fig. 3l–n, s). The effects of A > C, A > T, and A > G induced by the separate components of CC-AYBE were minimal (Supplementary Fig. 7a–c). The undesired indel byproduct generated by CC-AYBE was generally either lower than or comparable to that of intact AYBE at all loci (Fig. 3o). CC-AYBE exhibited editing efficiency for A-to-Y transversion at A₇ and A₈ that was similar to or slightly lower than that of intact AYBE (Fig. 3p–r). Taken together, these results demonstrate that coiled-coil heterodimers can be widely adapted to various base editing tools, significantly enhancing flexibility.

Engineering the CC-BE system to enhance its targeting scope

To broaden the targeting scope of CC-BE system, we developed NG-CC-BE (NG-CC-BE3, NG-CC-ABE8e) using SpCas9 variant (SpCas9-NG), which had been engineered to recognize relaxed NG PAM⁵¹. We tested NG-CC-BE3 (NG-sCBE-P3P4, NG-sCBE-N5N6) and NG-CC-ABE8e (NG-sABE-P3P4, NG-sABE-N5N6) in HEK293T cells at three endogenous genomic sites with NG PAM for each editor. The results showed that NG-CC-BE3 was substantially better than NG-BE3-ctrl, NG-intein-BE3 (Npu C574, Npu S714), and intact NG-BE3 at the tested loci, accordingly obtained slightly higher indel rates at SHANK3 (Fig. 4a–e). Notably, the A-to-G conversion level of NG-sABE-P3P4 was similar to or slightly higher than that of the intact NG-ABE8e and NG-intein-ABE8e (Npu C574, Npu S714), and more efficient than NG-sABE8e-ctrl, while the indel rate of NG-sABE-P3P4 was comparable to or lower than that of intact NG-ABE8e at all tested loci (Fig. 4f–j). We should recognize that a broader editing window may increase the risk of bystander editing, necessitating careful evaluation of the suitability of selected gRNAs.

Fig. 4. University of CC-BE system reflecting in its expanded editing scope and application in primary cells.

a–d The C-to-T editing efficiency of NG-sCBE-P3P4, NG-sCBE-N5N6, NG-sCBE-ctrl, NG-CBE, and split-intein NG-CBE (Npu C574, Npu S714) was examined at 3 loci in HEK293T cells. Data represent mean ± standard deviation of 4 independent biological replicates. e The undesired indel efficiencies of NG-sCBE-P3P4, NG-sCBE-N5N6, NG-sCBE-ctrl, NG-CBE, and split-intein NG-CBE (Npu C574, Npu S714) at 3 loci in b–d. Data represent mean ± standard deviation of 4 independent biological replicates. f–i The A-to-G editing efficiency of NG-sABE-P3P4, NG-sABE-N5N6, NG-sABE-ctrl, NG-ABE, and split-intein NG-ABE (Npu C574, Npu S714) was examined at 3 loci in HEK293T cells. Data represent mean ± standard deviation of 4 independent biological replicates. j The undesired indel efficiencies of NG-sABE-P3P4, NG-sABE-N5N6, NG-sABE-ctrl, NG-ABE8e, and split-intein NG-ABE (Npu C574, Npu S714) at 3 loci in g–i. Data represent mean ± standard deviation of 4 independent biological replicates. k–m Target C-to-T base editing (k, l) and undesired indel efficiencies (m) of sCBE-P3P4, sCBE-N5N6, sCBE-ctrl, and unsplit CBE system in bulk PFF cells across 4 endogenous loci. Data represent mean ± standard deviation of 3 independent biological replicates. n–p Target A-to-G base editing (n, o) and undesired indel efficiencies (p) of sABE-P3P4, sABE-N5N6, sABE-ctrl, and unsplit ABE system in bulk PFF cells across 4 endogenous loci. Data represent mean ± standard deviation of 3 independent biological replicates. Statistical significance was determined via two-tailed Student’s t tests. Exact P values are provided in Source data file. (*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, ns indicating not significant). Source data are provided as a Source data file.

These data demonstrate that the CC-BE system is broadly compatible with previously characterized SpCas9 variants, thereby broadening the targeting scope and potential applications of the CC-BE through enhanced PAM flexibility.

Robust base editing with CC-BE in primary porcine fetal fibroblasts

To further authenticate the editing activity of CC-BE in other mammalian cells, we selected 4 endogenous sites (GHR, HMGA2, IGF1, and PUM2) for CC-CBE and 4 endogenous sites (GHR, IGF1, PUM2, and SMAD2) for CC-ABE in porcine fetal fibroblasts (PFF) cells. These genes were previously reported to affect body size in mice, dogs, and pigs^52–54. We electrotransfected PFF cells with site-specific sgRNAs and CC-BE expression plasmids. The electrotransfection efficiency, measured using a reporter system, was approximately 80% (Supplementary Fig. 8a). Genomic DNA of the transfected bulk cells was extracted 72 h post-transfection, and targeted deep sequencing was employed to determine the editing efficiency.

According to the results, the CC-CBE demonstrated substantially improved editing efficiencies compared to the unsplit CBE system with peak activity at 4–7 positions (Fig. 4k, l). Strikingly, the increased editing efficiency of CC-CBE was most evident at the IGF1 site with a 12.4-fold change. The sCBE-ctrl exhibited unstable efficiencies in comparation to the unsplit CBE, but still lower than CC-CBE, which was consistent with the data above (Fig. 4k). Compared to unsplit CBE system, the CC-CBE induced lower or similar unwanted genomic alterations at 2 loci (Fig. 4m). Comparable base editing efficiencies were attained by the sABE-P3P4 and unsplit ABE system (Fig. 4n). And the sABE-N5N6 exhibited somewhat lower editing efficiencies (Fig. 4n). In contrast, the direct split ABE system exhibited substantially lower efficiencies compared to CC-ABE and the unsplit ABE system across all 4 edits in PFF cells (Fig. 4n). Both CC-ABE and unsplit ABE system demonstrated high efficiencies with observed peak activity at 4-8 positions (Fig. 4o). The indel frequencies of both CC-ABE and the directly split ABE were lower than those of the unsplit ABE (Fig. 4p). Computing an average editing efficiency at each position, the results showed that editing window was much wider when base editors were divided into two parts by coiled-coil heterodimers (Fig. 4l, o). Thus, CC-BE systems were able to efficiently edit positions that were previously challenging to target with canonical BE systems in PFF cells. In conclusion, our findings highlight that CC-ABE and CC-CBE achieve promising base editing efficiencies in PFF cells, suggesting their capability for mammalian gene editing and potential for broad applications.

Assessing off-target effects of CC-BE

Noting the high degree of CC-BE for on-target editing, we next considered the editing fidelities of CC-BE. BEs may bind to the off-target site that are similar to the target sgRNA protospacer, leading to sgRNA-dependent off-targeting. For the assessment of these rates, several off-target sites within 3 nucleotide mismatches have been identified for unsplit base editors, direct split base editors, and CC-BEs via in silico prediction (Supplementary Fig. 9a). The on- and off-target activities of these base editor effectors were measured by deep sequencing. The split and unsplit systems showed similar performance at Chr4 site and EMX1 site 4 (Supplementary Fig. 9b, c). At HIRA site 1, both sCBE-P3P4 and sCBE-N5N6 moderately increased the rates of sgRNA-dependent off-targeting (Supplementary Fig. 9b). And at Chr15 site, the sABE-N5N6 showed slightly lower off-target efficiencies than intact ABE system, while the sABE-P3P4 exhibited somewhat higher off-target efficiencies (Supplementary Fig. 9c). It is worth noting that the sgRNA-dependent off-target generated by CC-BE was not evidently increased compared to the untreated groups in general (Supplementary Fig. 9b). The off-target rate of CC-CBE was below 0.6% for all tested alleles, and the CC-ABE typically exhibited off-target efficiencies below 0.5%, except for sABE-P3P4 at Chr15 site OT-1 (Supplementary Fig. 9b, c).

Besides sgRNA-dependent off-target base editing, the deamination that occurs when the deaminase domain of the base editor binds to DNA without the involvement of sgRNA represents a distinct form of off-target base editing^55,56. Thus, we turned our attention towards sgRNA-independent off-target editing. We used the orthogonal R-loop assay⁴⁸ with dead SaCas9 to generate off-target R-loops and capture the guide-independent DNA editing mediated by CC-BE, direct split BE, and unsplit BE systems (Supplementary Fig. 10a). Amplicon deep sequencing was used to evaluate the on- and off-target activities of these base editors, focusing on the guide-target sites as well as three R-loops generated by dSaCas9. Using this system, off-target sgRNA-independent deamination at the dSaCas9 R-loop site was readily detectable for sCBE-P3P4, but the ratio of on-target to off-target editing for sCBE-P3P4 was higher than intact BE3 at 2 loci out of 3 edits for Sa site1(Supplementary Fig. 10b). Despite having a lower on-target to off-target editing ratio than the intact CBE at other Sa sites, sCBE-P3P4 has much higher precision editing efficiencies (Supplementary Fig. 10b and Fig. 1e). For sCBE-N5N6, off-target activity was lower than or comparable to that of intact BE3 at Sa site1, which consequently led to a higher ratio of on-target to off-target editing (Supplementary Fig. 10b). Notwithstanding that sCBE-N5N6 showed a lower ratio of on-target to off-target editing at Sa site6, a notable improvement was achieved in base editing efficiency (Supplementary Fig. 10b and Fig. 1e). The sABE-P3P4 exhibited notable off-target activity compared to the other groups, excluding the direct split ABE system, across all three off-target sites, implying that systems with high on-target efficiency also tend to have high off-target effects (Supplementary Fig. 10c and Fig. 2e). We also found that sABE-N5N6 system was less prone to editing the orthogonal R-loops so that they outperformed others for the on-target: off-target editing ratios at all detected sites (Supplementary Fig. 10c). These results demonstrated that sBE-N5N6 exhibited high targeting specificity. Based on all of the results above, we observed that CC-BE systems offered two options for genetic therapies: high specificity of sBE-N5N6 and high efficiency of sBE-P3P4.

To investigate the off-target effects of CC-BE at the genome-wide level, we performed whole-genome sequencing (WGS) in HEK293T cells. A total of 6 groups were incorporated: 3 CBE groups with or without sgRNA of EMX1 site4 (sCBE-P3P4, sCBE-N5N6, and unsplit CBE), 3 ABE groups with or without sgRNA of Chr4 site (sABE-P3P4, sABE-N5N6, and unsplit ABE). We transfected plasmids into HEK293T cells. Bulk cells were harvested 72 h post-transfection, and 30× WGS was performed. First, we confirmed the on-target base editing efficiency by amplicon deep sequencing (Supplementary Fig. 11). Then, genome-wide single-nucleotide variants (SNVs) and small insertions or deletions (indels) were analyzed based on WGS data in CC-BE3 (sCBE-P3P4, sCBE-N5N6), CC-ABE8e (sABE-P3P4, sABE-N5N6), unsplit BEs (BE3, ABE8e), and the non-targeting groups (with base editors but without sgRNAs, NT). The results revealed comparable levels of SNVs (Fig. 5a, b) and indels (Fig. 5c, d) between CC-BE and unsplit BE, as well as with sgRNA-treated and NT groups. Among the CBE groups, there was no significant difference in the average numbers of C > T SNVs in sCBE-P3P4, sCBE-N5N6, CBE, sCBE-P3P4-NT, sCBE-N5N6-NT, and CBE-NT (Fig. 5e). For the ABE groups (sABE-P3P4, sABE-N5N6, ABE, sABE-P3P4-NT, sABE-N5N6-NT, and ABE-NT), the average numbers of A > G SNVs were similar in each group (Fig. 5f). The base substitution types of SNVs (Supplementary Fig. 12a, b) and lengths of indels (Supplementary Fig. 12c, d) showed almost no bias in each group. We mapped the distribution of SNVs to the genome and found that there were almost no differences among the samples (Fig. 5g, h). For the exonic regions, minimal numbers of C > T SNVs in CBE groups and A > G SNVs in ABE groups were detected, and showed concordance among samples (Fig. 5i, j). The genomic distribution of all SNVs, C > T SNVs in CBE groups and A > G SNVs in ABE groups, revealed by circos plots contributed throughout the genome, and there were no regional and sample biases (Fig. 5k, l). In summary, our results suggest the off-target effects of CC-BE are minimal and similar to unsplit BE.

Fig. 5. Whole genome sequencing analysis of CC-BE in HEK293T cells.

a–d Cumulative (left) and normalized (right) frequency distributions of SNVs (a) and indels (c) in sCBE-P3P4, sCBE-N5N6, CBE, sCBE-P3P4-NT, sCBE-N5N6-NT, and CBE-NT. Cumulative (left) and normalized (right) frequency distributions of SNVs (b) and indels (d) in sABE-P3P4, sABE-N5N6, ABE, sABE-P3P4-NT, sABE-N5N6-NT, and ABE-NT. e, f The numbers of C > T SNVs in CBE groups (e) and A > G SNVs in ABE groups (f). g, h Bar chart of genome distribution of SNVs in CBE (g) and ABE (h) groups. i, j The numbers of C > T SNVs in CBE groups (i) and A > G SNVs in ABE groups (j) at exionic region. k, l Circos plot of the distribution of all SNVs and C > T SNVs in CBE groups (k), all SNVs and A > G SNVs in ABE groups (l). NT nontarget group. Source data are provided as a Source data file.

CC-BE-mediated correction of Fah point mutation

To explore the therapeutic potential of CC-ABE, we selected the mutant mouse fumarylacetoacetate hydrolase (Fah) gene for validation. Defects in this gene cause hereditary tyrosinemia type I (HTI), a lethal genetic metabolic disorder. The Fah gene encodes a key enzyme in the tyrosine metabolism pathway, and deficiency in this gene leads to the accumulation of toxic metabolic intermediates, resulting in hepatocyte apoptosis and severe liver damage⁵⁷. We evaluated CC-BE in Fah mutant mouse embryonic fibroblasts (MEFs) carrying a homozygous G-to-A point mutation at the end of exon 8, which induces exon skipping and results in the loss of functional Fah enzyme (Supplementary Fig. 13a). To achieve precise base-editing using ABE, we utilized a previously validated Fah sgRNA for targeting⁵⁸. Within the base editing window of the sgRNA, there are two adenines located at positions 6 and 9, and the adenine at position 9 (A₉) is the pathogenic mutation site. The A-to-G base conversion of bystander A₆ does not restore splicing. The MEFs were co-electroporated with Fah-targeting sgRNA and the corresponding nucleases. The transfection efficiency is approximately 50% (Supplementary Fig. 8b). Sanger sequencing revealed that sABE-P3P4 mediated evident A-to-G base editing at position A₉, resembling unsplit ABE (Supplementary Fig. 13b). Consistent with the aforementioned results, deep sequencing of the transfected bulk cells confirmed that sABE-P3P4 efficiently converted A₉-to-G (Supplementary Fig. 13c). The average conversion efficiencies of sABE-P3P4 were comparable to those of unsplit ABE (14.9% vs 16.4%), demonstrating therapeutic potential (Supplementary Fig. 13c).

CC-BE enables efficient in vivo targeting of Pcsk9 gene by dual-AAV

To validate the in vivo application of CC-ABE, we selected murine proprotein convertase subtilisin/kexin type 9 (Pcsk9) gene, a candidate target for base editing in the treatment of familial hypercholesterolemia, as its loss leads to a reduction in low-density lipoprotein (LDL) cholesterol⁵⁹. The sgRNA targets the splice donor site of Pcsk9 intron 1, within which its base editing window includes two adenines located at positions 4 (A₄) and 6 (A₆) (Fig. 6a). The desired editing involves A₆ to G conversion, leading to retention and read-through of intron 1, thereby disrupting Pcsk9, while the bystander A₄ is a nonfunctional base within the intron (Fig. 6a). First, we evaluated the base editing efficiency of CC-ABE at Pcsk9 locus in bulk MEFs in vitro. We observed a mean of 45.6% targeted A₆ editing with sABE-P3P4, while unsplit ABE was 38.7% (Supplementary Fig. 13d). Considering that the A-to-G conversion efficiencies of sABE-P3P4 were higher than those of sABE-N5N6, we packaged sABE-P3P4 into dual-AAV vectors (Supplementary Fig. 13d). The first vector carried TadA8e tethered with coiled-coil peptide P3 at the C-terminus driven by EF-1a short (EFS) promoter and the sgRNA targeting Pcsk9 driven by U6 promoter (Supplementary Fig. 13e). The second vector carried nCas9(D10A) tethered with coiled-coil peptide P4 at the N-terminus driven by EFS promoter (Supplementary Fig. 13e). The direct split ABE system without tethers was packaged into AAV as the control (Supplementary Fig. 13e). All four vectors were packaged into AAV serotype 8, which exhibits high tropism for hepatocytes. We delivered the dual-AAV8 sABE-P3P4 or sABE-ctrl (5 × 10¹¹ viral genomes each) to 1-week-old mice via intraperitoneal injection (Supplementary Fig. 13f). Two weeks later, twelve pieces of liver samples were collected from each mouse, and genomic DNA was isolated for evaluating base editing efficiency. Notably, we obtained unambiguous Sanger sequencing traces for A₆-to-G conversion events mediated by sABE-P3P4 compared to sABE-ctrl (Supplementary Fig. 13g). Consistently, high-throughput sequencing showed that the mean of targeted A₆ A-to-G conversion mediated by sABE-P3P4 was 20.4%, reaching as high as 33.84%, significantly higher than that of sABE-ctrl (Supplementary Fig. 13h, i). The average indel efficiencies of sABE-P3P4 (n = 6) and sABE-ctrl (n = 5) were 0.09% and 0.03%, respectively (Supplementary Fig. 13j, k).

Fig. 6. Dual AAV-mediated base editing of the mPcsk9 gene in vivo.

a Schematic of ABE-mediated mPcsk9 gene disruption. The PAM sequence is marked in green, and the adenines within editing windows are highlighted in red. b Schematic of split Cas9-based sABE-P3P4 and sABE-ctrl. c Experimental design for assessing the effect of mPcsk9 base editing in vivo. The 4-week-old mice were administered AAV8 via intraperitoneal injection (high-dose group: sABE-P3P4, n = 2; sABE-ctrl, n = 2; low-dose group: sABE-P3P4, n = 3; sABE-ctrl, n = 3) or intravenous injection (high-dose group: sABE-P3P4, n = 3; sABE-ctrl, n = 3; low-dose group: sABE-P3P4, n = 3; sABE-ctrl, n = 3). Liver tissues and serum were harvested 6 weeks post injection. d Sanger-sequencing chromatograms of sABE-P3P4 and sABE-ctrl show A-to-G base conversion (marked as T-to-C conversion) at the Pcsk9 gene in AAV-treated mice injected intraperitoneally. e, f The target A-to-G base conversion frequencies were measured by deep sequencing in the bulk livers of sABE-P3P4 and sABE-ctrl, 6 weeks after AAV8 injection. g–k Quantification of Pcsk9 (g), total cholesterol (h), LDL cholesterol (i), AST (j), and ALT (k) levels in AAV-injected and PBS-injected mice. LDL low-density lipoprotein, AST aspartate transaminase, ALT alanine aminotransferase. Source data are provided as a Source data file.

Next, we systematically compared editing frequencies in vivo among different injection approaches and AAV doses using split Cas9-based CC-ABE (Fig. 6b). Four-week-old mice were injected intraperitoneally or intravenously with different doses of AAV8 (low-dose condition: 2.5 × 10¹¹ viral genomes each; high-dose condition: 5 × 10¹¹ viral genomes each) (Fig. 6c). Six weeks after injection, liver tissues and serum were harvested to assess genome-editing events (Fig. 6c). Sanger sequencing demonstrated strong performance for A₆-to-G conversion in the liver injected intraperitoneally with sABE-P3P4, but not with sABE-ctrl under high-dose conditions (Fig. 6d). Deep sequencing of amplicons indicated high frequencies of A₆-to-G conversion in the sABE-P3P4-treated groups under both high-dose (~70%) and low-dose conditions (~50%) (Fig. 6e, f). The editing efficiency of intraperitoneal injection (n = 2 for the high-dose group and n = 3 for the low-dose group) and intravenous injection (n = 3 for each group) was comparable, and the efficacy exhibited a dose-dependent manner (Fig. 6e, f). Livers treated with sABE-P3P4 showed much higher editing efficiencies compared to those treated with sABE-ctrl (Fig. 6f). We assessed whether bystander edits could be generated by CC-ABE at this site. As shown in Fig. 6f and Supplementary Fig. 14, bystander edits of A₄ were observed, while they occurred in the intronic region and would not impact the transcription of the gene. Furthermore, we demonstrated the functionality of base editing. The levels of Pcsk9, total cholesterol, and LDL cholesterol in the sABE-P3P4-treated mice were decreased compared to those in the PBS-treated controls (Fig. 6g, h, i). The levels of liver function markers, alanine aminotransferase (ALT) and aspartate transaminase (AST), in the AAV-injected mice were not elevated, indicating no acute liver dysfunction (Fig. 6j, k).

Taken together, these data strongly demonstrate that CC-BE can be successfully delivered and achieve base editing in vivo with high efficiency.

Delivery of NG-CC-ABE via dual-AAV9 restores dystrophin expression in ∆Ex51 mice

Duchenne muscular dystrophy (DMD) is a severe monogenic muscle disorder resulting from mutations in the X-linked Dmd gene, which encodes dystrophin, and affects approximately 1 in 3500–5000 newborn males⁶⁰. The absence of dystrophin results in membrane fragility and muscle degeneration, leading to walking incapability during adolescence and premature death from cardiac and respiratory failure in early adulthood. Becker muscular dystrophy (BMD) is categorized as a relatively mild muscle degenerative disorder resulting from in-frame deletions of the Dmd gene, supporting the synthesis of dystrophin proteins with partial functionality⁶⁰. To verify the ability of CC-BE to correct pathogenic variants in non-liver tissue disease, for which the lipid nanoparticles (LNPs) would not be an effective alternative therapy, we employed a dual-AAV-based CC-BE strategy to convert DMD to BMD, aiming to restore dystrophin expression in male mice harboring a deletion of exon 51 (∆Ex51) of the Dmd gene (Supplementary Fig. 15), which represents a prevalent disease-causing mutation in humans⁶¹.

Proper exon splicing requires 3′ splice acceptor sites (SAS) and 5′ splice donor sites (SDS). To correct the dystrophin reading frame in male ∆Ex51 mice (Supplementary Fig. 15), we used the sgRNA (mEx50-sgRNA)⁶² with NG PAM, which induces exon skipping by disrupting the SDS of exon 50 via NG-ABE-mediated gene targeting (Fig. 7a). First, we evaluated the editing efficiency of NG-CC-ABE8e, NG-ABE8e-ctrl, and unsplit NG-ABE8e in mouse N2a neuroblastoma cells. Consistent with the data above, NG-CC-ABE8e exhibited A-to-G base editing efficiency comparable to that of unsplit NG-ABE8e, and obviously higher than that of NG-ABE8e-ctrl at the SDS of exon 50 (Fig. 7b). NG-sABE8e-P3P4, with NG-sABE8e-ctrl as a control, was chosen for in vivo base editing. AAV vectors were constructed, with each vector containing an EFS-driven half of NG-sABE8e-P3P4 or NG-sABE8e-ctrl and a U6-driven mEx50-sgRNA in the reverse orientation (Fig. 7c). All vectors were packaged into AAV9, which displays preferential tropism for skeletal and cardiac muscles⁶³. Dual-AAV split NG-ABE8e system was initially delivered to 5-week-old male ∆Ex51 mice via local intramuscular injection (2.5 × 10¹¹ viral genomes each) or intraperitoneal injection (5 × 10¹¹ viral genomes each), with two mice per group (Fig. 7d). Muscles were collected for analysis at 7 weeks post-injection (Fig. 7d). For the intraperitoneally injected mice, the tibialis anterior (TA), gastrocnemius (GCM), diaphragm (DIA), cardiac muscle (CM), and abdominal muscle (ABS) were harvested; for those receiving local intramuscular injection, only the TA and GCM were collected. Four pieces of samples were harvested from each muscle of AAV9-injected male ∆Ex51 mice, PBS-injected male ∆Ex51 mice, and wild-type mice. Genomic PCR amplification products spanning the target site were subjected to amplicon deep sequencing. In intraperitoneally injected mice, the CM (ranging from 7.8% to 15.6%), GCM (ranging from 2.7% to 5.1%), ABS (ranging from 1.8% to 4.8%), TA (ranging from 1.7% to 3.6%), and DIA (ranging from 0.5% to 1.4%, with one notably higher value of 6.0%) exhibited efficient base editing (Fig. 7e). In mice receiving local intramuscular injection, the efficiencies of GCM ranged from 6.1% to 22.2%, and TA ranged from 4.0% to 8.9% (Fig. 7e). Local injection was more effective than intraperitoneal injection in the TA and GCM (Fig. 7e, f). Notably, regardless of whether the injection was intraperitoneal or local intramuscular, the editing efficiencies of NG-sABE8e-P3P4 surpassed those of NG-sABE8e-ctrl (ranging from 0% to 1.3%) in every muscle type tested (Fig. 7e, f). RT-PCR products derived from the TA, GCM, and CM of male ∆Ex51 mice injected with the NG-sABE8e-P3P4 system revealed exon 50-skipping events (Fig. 7g, lower band). These exon 50-skipping events were further confirmed by Sanger sequencing of the reverse transcription products, which demonstrated that exon 49 was spliced to exon 52, restoring the reading frame of the Dmd transcript (Fig. 7h).

Fig. 7. Exon skipping by dual AAV9-based NG-CC-ABE in the ∆Ex51 mice.

a Schematic showing exon skipping strategies to restore the ORF of the Dmd transcript. b On-target DNA editing efficiencies in mouse bulk N2a cells after transfection with NG-CC-ABE, NG-ABE-ctrl, and unsplit NG-ABE. Data represent mean ± standard deviation of 4 independent biological replicates. c Schematic of the dual-AAV9 system for in vivo delivery of NG-sABE and mE50-sgRNA. d Overview of the in vivo injection of dual-AAV9 system in ∆Ex51 mice. e, f On-target base editing efficiencies of NG-sABE-P3P4 and NG-sABE-ctrl in bulk muscles at 7 weeks post-AAV9 injection. (Intraperitoneal injection: NG-sABE-P3P4, n = 2; NG-sABE-ctrl, n = 2; local intramuscular injection: NG-sABE-P3P4, n = 2; NG-sABE-ctrl, n = 2). g RT-PCR analysis of RNA extracted from TA, GCM, and CM of wild-type mice, ∆Ex51 mice, and NG-sABE-P3P4-treated mice. Each experiment was repeated independently 2 times with similar results. h Sequence of RT-PCR products from wild-type mice, ∆Ex51 mice, and NG-sABE-P3P4-treated mice. TA tibialis anterior, GCM gastrocnemius, DIA diaphragm, CM cardiac muscle, ABS abdominal muscle. Source data are provided as a Source data file.

Based on the evident gene correction efficiency observed after AAV9 injection, we tested whether dystrophin expression levels were rescued in the TA, GCM, and CM extracted from AAV9-injected mice. Using laminin to label muscular tissue, immunostaining revealed that dystrophin expression was restored in mice injected with NG-sABE8e-P3P4, while it was barely detectable in those injected with NG-sABE8e-ctrl in the TA (Fig. 8a), GCM (Fig. 8b), and CM (Fig. 8c). Fully in line with the results above, dystrophin expression levels in the locally injected TA and GCM were higher than in those injected intraperitoneally (Fig. 8a, b). Additionally, the levels of serum creatine kinase (CK) and α-hydroxybutyrate dehydrogenase (α-HBDH), two important markers of muscle injury, were much closer to those of wild-type mice in those injected with NG-sABE8e-P3P4 (Fig. 8d, e). The levels of lactate dehydrogenase (LDH), ALT, and AST in the NG-sABE8e-P3P4-injected mice were not drastically elevated compared to those of wild-type mice, suggesting that gene therapy did not cause liver or other organ damage and may even have a mitigating effect (Fig. 8f, g, h).

Fig. 8. Dystrophin restoration following NG-CC-ABE-mediated base editing in the ∆Ex51 mice.

a–c Immunostaining for laminin, dystrophin, and DAPI in TA (a), GCM (b), and CM (c) of wild-type mice, AAV9-injected mice, and ∆Ex51 control mice. Immunostaining was performed 7 weeks after injection of AAV9. Scale bars, 100 µm. Each experiment was repeated independently 2 times with similar results. d–h Serum CK (d), α-HBDH (e), LDH (f), ALT (g), and AST (h) levels in different groups at 7 weeks post-AAV9 injection (n = 2 in each group). TA tibialis anterior, GCM gastrocnemius, CM cardiac muscle, CK creatine kinase, α-HBDH α-hydroxybutyrate dehydrogenase, LDH lactate dehydrogenase, ALT alanine aminotransferase, AST aspartate transaminase. Source data are provided as a Source data file.

Collectively, these data robustly demonstrate that CC-BE can be effectively delivered for precise in vivo editing in non-liver tissues, highlighting the therapeutic potential of the CC-BE system.

Discussion

The landmark creation of base editor has enabled genome manipulation through single-base substitution, holding promise for correcting pathogenic point mutations^3,4. The challenge of packaging base editor into AAV vectors has inspired the development of feasible technology capable of mediating base editing in vivo. Although several strategies have been developed, including intein-mediated split base editors and miniature base editors, these technologies are either labor-intensive or compromise base editing efficiency²⁷.

In the present study, we developed a platform for splitting base editor with coiled-coil dimerization peptides, termed CC-BE, in which nCas9 and deaminase bind to each other via CC-mediated electrostatic and hydrophobic interaction. Our developed CC-BE demonstrates higher or similar editing efficiency compared to the previously reported split-intein base editor. Additionally, the CC-BE system exhibits broad versatility, allowing for engineering multiple base editors, including ABE8e, BE3, TadCBE, ABE9, and AYBE, and enabling efficient base editing across diverse genomic loci, various cell types, and SpCas9 variants with extensive targeting scope. Interestingly, we observed that CC-ABE functions as effectively as intact ABE, while CC-CBE enhances the base editing efficiency compared to the unsplit CBE. Through 3D structure predictions using AlphaFold 3, we found that in the presence of coiled-coil dimerization peptides, the spatial positions of UGI and APOBEC3 are closer compared to the intact CBE. Therefore, we reasonably infer that the conformational change significantly alters the overall performance of CC-CBE, leading to an improvement in base editing efficiency. However, an obvious improvement is observed in CBE systems, but not in ABEs. This may be due to several reasons. First, the conformational changes induced by CC peptides may enhance the activity of both UGI and APOBEC3, but have no effect on adenine deaminase alone. Second, differences in the reaction kinetics and substrate-binding affinity of cytosine and adenine deaminases may lead to varying gene editing outcomes. In practical applications, we treated fibroblast from Fah mutant mice with CC-BE, successfully achieving gene correction. Furthermore, we conducted in vivo validation of the editing efficiency by delivering AAV-packaged CC-BE into mice for base editing at the Pcsk9 locus, resulting in high in vivo editing efficiency in the mouse liver (up to 79.0%). We also employed a dual-AAV-based CC-BE strategy to treat DMD mice harboring a deletion of exon 51. Our data suggest that single-swap genomic editing at the Dmd SDS using the dual-AAV-based CC-BE system is efficient in inducing exon 50 skipping, thereby restoring the open reading frame of the dystrophin transcript and rescuing dystrophin expression. This demonstrates the potential of CC-BE for in vivo therapeutic editing.

Compared to conventional BEs, CC-BEs have evident advantages, including flexible design, improved or comparable base editing efficiency, and the ability to overcome the AAV packaging issue caused by the large size of BEs. The in vivo delivery of CC-BE relies on dual-AAV system, whereas delivering miniature base editors based on hypercompact Cas proteins such as Cas12f, TnpB, and IscB requires single-AAV. However, these miniature gene editing tools demonstrate lower editing activity compared to conventional BEs and have a limited editing scope, fundamentally restricting their capacity for precise targeting. Although dual AAV delivery methods have been explored to address size limitations, these strategies introduce additional manufacturing and safety challenges due to the necessity of co-delivering multiple vectors. In the future, compact nucleases with potent editing capabilities, such as NanoCas variants⁶⁴, may enable single-AAV delivery of effective base editors in vivo, thus combining the advantages of high efficiency and streamlined delivery.

In our assessment of CC-BE off-target effects, we observed variability in off-target efficiency depending on the site and variant, yet consistently within acceptable limits. Our WGS results demonstrated that the off-target effects of CC-BE were minimal and comparable to those of unsplit BE. The subsequently developed higher-fidelity deaminases can be utilized for the optimization of CC-BE system.

Comprised of just 28 amino acids, coiled-coil peptides serve as a flexible tool for facilitating protein-protein interactions, mediated through hydrophobic and electrostatic interactions^45,46. Coiled-coil peptides have been employed in engineering genome editing tools, such as recruiting an exonuclease to Cas9 to enhance gene editing efficiency⁶⁵. Notably, our research group has devised a flexible strategy for splitting prime editor via coiled-coil peptides, enabling highly efficient prime editing both in vitro and in vivo⁶⁶. Here, we creatively expand the coiled-coil-based gene editing toolkit.

In summary, the CC-BE system developed in this study brings several strengths, including simple design with dimer-forming modules, high base editing efficiency, and compatibility across various base editors, cell types, and Cas9 variants. The CC-BE system provides a more straightforward and effective strategy for base editing, particularly facilitating the utility of base editors in vivo.

Methods

Mice

For targeting the mPcsk9 gene, experiments were conducted using 16 wild-type (C57BL/6J) mice aged 1 week and 26 wild-type (C57BL/6J) mice aged 4 weeks. When targeting the mPcsk9 gene, sex was not considered because this gene is located on the autosome. Given that Duchenne muscular dystrophy (DMD) is a severe X-linked genetic disorder, only male mice were used for targeting the mDMD gene in vivo. For targeting the mDMD gene, 6 male ∆Ex51 mice (C57BL/6JGpt-Dmd^em2Cd1406/Gpt) aged 5 weeks were used in the experiments. The mouse housing facilities were kept at temperatures between 20–22 °C. The mice were housed in an environment with a 12-h light-dark cycle, provided with continuous access to standard rodent feed and drinking water.

Plasmid construction

The CC-ABE and CC-CBE editors used in this manuscript were constructed by pEASY-Uni Seamless Cloning and Assembly Kit (TransGen Biotech, CU101-02) using the pCMV-BE3/pCMV-ABE8e plasmid as the backbone. The P3, P4, N5, and N6 sequences (Supplementary Data 1) were integrated into the vector by primer synthesis. To facilitate enrichment of transfected cells in subsequent experiments, we cloned PGK-mCherry into the expression vectors of nCas9 and nCas9-UGI, and PGK-EGFP into the expression vectors of TadA8e, APOBEC and unsplit BEs, respectively.

To construct CC-ABE9, N108Q/L145T mutations were introduced into pCMV-TadA8e vector. To construct Cas9-split CC-ABE, CC-TadCBE, and CC-AYBE, nCas9 was split into two parts before Cys 574. The fragments of T1.46 and MPG sequence were derived from Addgene #193284 and #193967 via PCR. For CC-AYBE, TadA8e-nCas9 (amino acid 1–573)-P3/N5 and P4/N6-nCas9 (amino acid 574–713)-MPG were assembled into the CMV backbone. The CC-TadCBE plasmid was constructed by replacing TadA8e and MPG with T1.46 and 2x UGI, respectively.

Plasmids expressing dSaCas9 and U6-SaCas9 guide RNA were derived from PX601, in which D10A and N580A were induced via mutagenesis strategy, and then guide RNA cassette was inserted through BsaI digestion strategy. The SaCas9 guide RNA and SpCas9 guide RNA used in this study were listed in Supplementary Data 2.

Cell line authentication

The HEK293T (CRL-3216) and mouse N2a neuroblastoma (CCL-131) cell lines used in this study were obtained from the American Type Culture Collection (ATCC). These cell lines were authenticated by ATCC. PFFs were isolated from wild-type E35 pig fetuses, and MEFs were isolated from E13.5 wild-type or HT1 mouse embryos. These cells were validated by our laboratory.

Cell culture and transfection

HEK293T and N2a cells were cultured in high-glucose Dulbecco’s modified Eagle’s medium (DMEM; HyClone) supplemented with 10% fetal bovine serum (FBS; Gibco), and the PFFs or MEFs were cultured in DMEM supplemented with 15% FBS (Gibco), 1% nonessential amino acids (Gibco), 2 mM GlutaMAX (Gibco), and 1 mM sodium pyruvate (Gibco) at 37 °C and 5% CO₂. The medium was changed daily. Cells were passaged and expanded using 0.25% trypsin once the cell confluency reached approximately 90%.

For HEK293T cells, cells were digested into single cells and seeded into a 24-well plate before transfection. PEI transfection was performed when the cell confluency reached 40–60%. Prior to transfection, the medium is replaced with 2% FBS medium. PEI (Sigma, 408727; final concentration: 1 μg/μL) was gently mixed with 2% FBS medium and incubated at room temperature for 5 min. For each well, 3 μg of PEI and 25 μL of 2% FBS medium were mixed, 1 μg of plasmid (250 ng of sgRNA + 750 ng of base editor or 375 ng of sBE at each), and 25 μL of 2% FBS medium were mixed. After that, the PEI mixture was added to the plasmid mixture dropwise, gently stirred with a pipette, and allowed to sit at room temperature for 20 min. The HEK293T cells were then slowly added to the PEI and plasmid mixture for transfection. 12 h after transfection, cells were cultured in medium containing 10% FBS.

For PFFs and MEFs, cells were digested from 10 cm dishes, washed once with PBS, and then divided into 25 aliquots. Each aliquot was mixed with plasmid (1 μg of gRNA and 3 μg of BE or sBE [1.5 µg+ 1.5 μg]) in 110 µL PBS, followed by co-electrotransfection at 1350 V, 30 ms, 1 pulse by using the Neon™ transfection system (Life Technology). The medium was replaced 12 h after transfection, and the bulk cells were harvested 72 h post-transfection.

For N2a cells, Lipo transfection was performed when cell confluency reached 40%–60% in 24-well plates. According to the manufacturer’s protocols, 0.8 µL Lipo 8000^TM (Beyotime, C0533) and 500 ng plasmid (125 ng of sgRNA and 375 ng of BE or sBE) were mixed in 25 µL medium containing 2% FBS prior to transfection. The mixture was then added into N2a cells following medium replacement. Cells were cultured in medium containing 10% FBS after 12 h and harvested 72 h after transfection.

FACS, cell lysis, and PCR identification

MoFlo Astrios was used to perform fluorescence-activated cell sorting 72 h after transfection. After washing with PBS and digesting with 0.25% trypsin, cells were collected by centrifugation at 200 × g for 3 min and then resuspended in PBS. Double-positive cells with mCherry and EGFP were sorted for CC-BEs, EGFP-positive cells were sorted for unsplit BE systems, indicating the successful transfection of CC-BEs or unsplit BE systems. Further analysis was performed on roughly 10,000–50,000 sorted cells. NP40 was used to lyse the sorted cells under the following conditions: 56 °C for 60 min and 96 °C for 10 min. Following that, 2×Hieff Canace® Plus PCR Master Mix (With Dye) (YEASEN, 10154ES03) was used to perform PCR amplification of the target site.

High‑throughput sequencing

Amplification primers were designed near the target locus for library preparation, with specific primers amplifying fragments of about 190 base pairs. PCR amplification was carried out in two rounds. First round, the primers listed in the Supplementary Data 2 were used for the amplification process. Using cell lysate as the template, amplification was carried out for 30 cycles in a 20 μL reaction volume. For the second round, primers with unique Illumina barcodes were used for amplification. Using the first round PCR product as the template, amplification was performed for 15 cycles in a 50 μL reaction volume. Following purification with the HiPure Gel Pure DNA Mini Kit (Magen, D2111-03) and quantification with the Equalbit™ dsDNA HS Assay Kit (Vazyme), the second round PCR products were prepared for deep sequencing on the Illumina HiSeq X platform (Annoroad Gene Technology Corporation). CRISPResso2 (http://crispresso.pinellolab.org/submission) was used to analyze the high-throughput sequencing data⁶⁷.

Protein extraction and Western blot analysis

For Western blot analysis, HEK293T cells seeded in 24-well plates were transfected with CC-BE (sABE-P3P4/sCBE-P3P4) and unsplit BE system (unsplit-ABE/CBE), then harvested 72 h later. Cell proteins were extracted using radioimmunoprecipitation analysis buffer (RIPA) supplemented with PMSF. The mixtures were lysed on ice for 5 min, and the supernatant was obtained by centrifugation at 15,000 × g for 15 min at 4 °C. Protein concentration was determined using the bicinchoninic acid (BCA) assay, and 60 µg of total protein was loaded onto an 8% acrylamide gel. Gels were run at 70 V for 150 min, followed by 120 min transfer to a polyvinylidene fluoride (PVDF) membrane at 200 mA on ice. The blot was incubated with rabbit anti-Cas9 antibody (JM11-55, HUABIO, ET1703-85), and beta-actin antibody (AF7018, Affinity) at 4 °C overnight, followed by incubation with HRP-conjugated Goat Anti-Rabbit IgG antibody (GB23303, Servicebio) at room temperature for 1 h. The blot was developed using Western Blotting Luminol Reagent (E411-05, Vazyme).

Off-target analysis

Potential off-target sites were predicted in the human genome with COSMID (https://crispr.bme.gatech.edu)⁶⁸. Three days after transfection with CC-BE and target sgRNAs (EMX1 site4 and HIRA site1 for CBE systems; Chr4 site and Chr15 site for ABE systems), cells were harvested for fluorescence-activated cell sorting. The sorted cells were then lysed with NP40. Sequences flanking the predicted off-target sites were amplified using primers designed by COSMID (Supplementary Data 3) for high-throughput sequencing.

R-loop assay

HEK293T cells were transfected at approximately 40%–60% confluency after being seeded on 24-well plates. 375 ng of intact BE systems or CC-BEs (187.5 + 187.5 ng), 250 ng of SpCas9 guide RNA plasmid, and 375 ng of dSaCas9 with SaCas9 guide RNA plasmid were co-transfected by PEI transfection. Twenty-four hours after transfection, the medium was replaced with medium containing 10% FBS. Three days post-transfection, cells were harvested for fluorescence-activated cell sorting. The sorted cells were lysed by NP40, and the PCR amplification was performed using primers flanking the SaCas9 guide-RNA target sites for high-throughput sequencing (Supplementary Data 4).

Whole genome sequencing data analysis

The WGS libraries and sequencing were prepared by Annoroad Gene Technology (Beijing, China). WGS was performed on Illumina NovaSeq X-25B at an average coverage of 30×. Fastp (v0.23.0) was used to filter raw reads (base quality value ≥ 19). The clean reads were mapped to the human reference genome (GRCh38 from UCSC) using BWA-MEM (v0.7.17). Samtools (v1.22) was used to sort BAM files. Picard (v2.9.0) was used to mark duplicative reads. Variants calling was conducted using GATK HaplotypeCaller (v4.1). Variants were annotated by ANNOVAR (v20220521). Barplots and heatmaps were generated using R package ggplot2 (v4.0.0). Genome-wide mutations were generated using 200 Kb sliding window and plotted with CIRCOS.

AAV production and injection

PackGene Biotech Co., Ltd. (Guangzhou, China) produced and packaged AAV8 and AAV9 viruses. Prior to injection, the mice were anesthetized with an abdominal injection of 5% chloral hydrate. For in vivo base editing of the mPcsk9 gene, we used AAV8 to package the viral vectors. For 1-week-old wild-type mice, dual-AAV8 sABE-P3P4 (5 × 10¹¹ genome copies of AAV-TadA8e-P3-sgRNA and 5 × 10¹¹ genome copies of AAV-P4-nCas9) or sABE-ctrl (5 × 10¹¹ genome copies of AAV-TadA8e-sgRNA and 5 × 10¹¹ genome copies of AAV-nCas9) were delivered via intraperitoneal injection. Two weeks after AAV injection, livers were harvested and divided into 12 pieces. For 4-week-old wild-type mice, injections were performed via intraperitoneal injection or intravenous injection with sABE-P3P4 (AAV-TadA8e-nCas9-N-P3-sgRNA and AAV-P4-nCas9-C-sgRNA) or sABE-ctrl (AAV-TadA8e-nCas9-N-Pcsk9-sgRNA and AAV-nCas9-C-Pcsk9-sgRNA) at different doses (low-dose condition: 2.5 × 10¹¹ viral genomes each; high-dose condition: 5 × 10¹¹ viral genomes each). After 6 weeks, the mice were euthanized, and liver tissues and serum were collected. TIANamp Genomic DNA Kit (TIANGEN) was used to obtain genomic DNA of liver samples.

For in vivo base editing of the mDmd gene, we used AAV9 to package the viral vectors. Dual AAV9 vectors of sABE-P3P4 or sABE-ctrl were delivered to 5-week-old ∆Ex51 mice (5 × 10¹¹ viral genomes of each AAV9 virus for intraperitoneal injection; 2.5 × 10¹¹ viral genomes of each AAV9 virus for local intramuscular injection). After 7 weeks, the TA, GCM, DIA, CM, and ABS were harvested from the intraperitoneally injected mice; only the TA and GCM were collected from those receiving local intramuscular injection. Each harvested muscle was divided into four parts, designated for subsequent analyses, including RNA extraction, DNA extraction, cryosectioning, and sample backup.

RT-PCR and Sanger sequencing

Muscle samples were ground by homogenizer (60 Hz, 60 s), and total RNA was extracted with RNA rapid extraction solution (Servicebio, G3013). The cDNA was PCR-amplified using the PrimeScriptTM RT reagent Kit with gDNA Eraser (Takara, RRO47A) according to the manufacturer’s protocol.

The cDNA isolated from WT mice, ∆Ex51 mice, and CC-BE-treated mice was subjected to PCR amplification using primers (listed in Supplementary Data 5) to obtain amplicons of 767, 534, and 425 bp size for Dmd Exon 48 to Exon 53. The PCR products were analyzed by 2% agarose gel electrophoresis and purified using the HiPure Gel Pure DNA Mini Kit (Magen, D2111-03) for Sanger sequencing.

Serum biochemistry analysis

Blood samples were collected from mice. The collected blood was allowed to stand at room temperature for 30 min before being centrifuged at 3000 rpm for 15 min to separate the serum. The samples were analyzed for various biochemical markers, including CK, ALT, AST, α-HBDH (HB3919), and LDH levels for wild-type mice, PBS-treated ∆Ex51 mice, and AAV9-injected ∆Ex51 mice. Additionally, Pcsk9, AST, ALT, total cholesterol, and LDL cholesterol levels were assessed in wild-type mice and AAV8-injected mice. Pcsk9 was measured using the Mouse PCSK9 (Proprotein convertase subtilisin/kexin type 9) QuickTest ELISA Kit (QT-EM0153), following the manufacturer’s instructions.

Immunofluorescence staining of muscles

Fresh muscle tissues were rapidly frozen using liquid nitrogen-cooled isopentane and stored at −80 °C for subsequent analysis. All muscles were embedded in OCT compound before being cryo-sectioned at 8 µm. The sections were fixed with acetone for 10 min and air-dried for 10 min at 4 °C. The sections were then stored at −20 °C.

Before immunofluorescence staining, the sections were rewarmed at room temperature for 15 min, followed by washing with PBS three times, 5 min each. After permeabilizing with 0.5% Triton X-100 for 30 min, the sections were blocked with 10% normal goat serum for 1 h at room temperature. After washing with PBS three times, the sections were incubated overnight at 4 °C with the antibody mixture (anti-laminin 2 alpha antibody[4H8-2], abcam, ab11576; dystrophin (PT0821R) PT® Rabbit mAb, Immunoway, YM8580). Each antibody was diluted at a ratio of 1:200. The next day, after washing 5 times, the sections were incubated with secondary antibodies (anti-rat IgG (H + L), CST, 4418; anti-rabbit IgG (H + L), CST, 4412) for 1 h at room temperature. Each secondary antibody was diluted at a ratio of 1:2000. DAPI mounting medium (Sigma-Aldrich, F6057-20ML) was added before applying the coverslip. Images were acquired using Zeiss LSM 800.

Statistics and reproducibility

In this study, data analysis was conducted using GraphPad Prism 10.3.1. Error bars indicated the mean ± standard deviation of at least three independent biological replicates. No statistical method was used to predetermine sample size. The sample size is based on the variability observed in independent experiments and is consistent with the standards in related studies. No data were excluded from the analyses. Investigators were not blinded to allocation during experiments or outcome assessment. For assessing significant differences in precise editing efficiency across groups, the two-tailed Student’s t test was employed (*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, and ns indicating not significant). Exact P values are provided in Source data file.

Ethics statement

C57BL/6J and ∆Ex51 mice (C57BL/6JGpt-Dmd^em2Cd1406/Gpt, GemPharmatech) were used in this study. All animal experiments were consistent with the guideline of the Animal Welfare and Research Ethics Committee at Guangzhou Institutes of Biomedicine and Health (GIBH), Chinese Academy of Sciences (IACUC: 2025073).

Reporting summary

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

Author contributions

H.S., K.W., L.L., and Q.J. designed this study. H.S., S.M., Q.L., and M.C. performed most of the experiments and analyzed the data. Q.L. performed the bioinformatic analysis. Z.L., Y.M., Y. Li, Y.S., and S.H. conducted part of the molecular experiments. Y.D., Y. Lin, and J.Z. carried out animal breeding and virus injection. J.J., Y.Y., X.R., and N.L. provided technical assistance. H.S., S.M., and Q.L. prepared the manuscript. H.S., K.W., L.L., and Q.J. supervised this study. All authors read and approved the final manuscript.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

All of the data used to support the study’s conclusions can be found in this article and the additional files. The amplicon sequencing and whole-genome sequencing data generated in this study have been deposited in the National Genomics Data Center, China National Center for Bioinformation under BioProject PRJCA033832. The raw data generated in this study are provided in the Supplementary Information/Source data file. Source data are provided with this paper.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-68469-2.