The DdCBE vectors used were synthesized in Sangon Biotech (Shanghai), which is composed of mitochondrial localization sequence (MTS), N terminal, C terminal, one kind of four split DddA halves (N terminal split at G1333 (G1333-N), C terminal split at G1333 (G1333-C), N terminal split at G1397 (G1397-N), and C terminal split at G1397 (G1397-C)), and UGI-coding sequences. Four pairs were designed for each site according to the two different splits with two different orientations (G1333CL+G1333NR, G1397CL+G1397NR, G1333NL+G1333CR, and G1397NL+G1397CR). G1333CL+G1333NR pair consists of G1333C fused to left TALE array (G1333CL) and G1333N fused to right TALE array (G1333NR); G1397CL+G1397NR pair consists of G1397C fused to left TALE array (G1397CL) and G1397N fused to right TALE array (G1397NR); G1333NL+G1333CR pair consists of G1333N fused to left TALE array (G1333NL) and G1333C fused to right TALE array (G1333CR); G1397NL+G1397CR pair consists of G1397N fused to left TALE array (G1397NL) and G1397C fused to right TALE array (G1397CR). TALE arrays were constructed using Golden Gate TALEN and TAL Effector Kit 2.0 (Addgene). The corresponding protein sequences are listed in the Supplementary Table S1. The Repeat Variable Diresidues (RVDs) containing NI, NG, NN and HD amino acids, recognize A, T, G and C, respectively. Ligated plasmids were transformed into Trans5α chemically competent cells (TransGene Biotech) and subjected to Sanger sequencing to analyze the identity of the constructs (Sangon Biotech). Final plasmids were prepared using Endofree mini plasmid kit II (TianGen) for cell transfection.

Neuro-2a cells (CCL-131; ATCC) were cultured in MEM (BasalMedia) with 10% fetal bovine serum (FBS, Gibco), 1% sodium pyruvate (Introvigen), 1% Non-Essential Amino Acids (NEAA, Gibco) and 1% penicillin–streptomycin (Gibco) at 37°C with 5% CO₂. U2-OS cells (HTB-96; ATCC) were cultured in high-glucose DMEM (Gibco)+10% FBS+1% penicillin–streptomycin at 37°C with 5% CO₂.

For transfection, Neuro-2a cells were plated in 24-well cell culture plates at a density of 2.4 × 10⁵ per well. After 24 h, lipofection was performed at a cell density of approximately 70%. Cells were transfected with 800 ng plasmid of each mitoTALE monomer. 0.75 μL LipofectamineTM 3000 Reagent and 1 μL P3000TM Reagent were used per well. 24 h after transfection, the original culture medium was replaced with a medium containing 1 mg/mL G418 (MCE). Cells were collected after screening for 72 h. U2-OS cells were plated in 12-well cell culture plates at a density of 3 × 10⁵ per well 24 h before lipofection. Cells were transfected with 1,600 ng plasmid of each mitoTALE monomer, using 1.5 μL LipofectamineTM 3000 Reagent and 2 μL P3000TM Reagent per well. G418 (1 mg/mL) screening starts 24 h after transfection and lasts for 5.5 days. Genomic DNA was extracted using the DNeasy Blood and Tissue Kit (Qiagen) and stored at −20°C till sequence library construction. 

Sanger sequencing was performed by Sangon Biotech (Shanghai). The primers for each spacer are listed in Supplementary Table S2. For NGS, Phanta Flash DNA Polymerase (Vazyme) was used to amplify target sequences with primers containing barcodes and Illumina adapters in the first round PCR (100 ng genome DNA extracted from cells as template). The sequence of primers is listed in Supplementary Table S2. PCR product of the first round was used for the second round of PCR using index primers (Vazyme). After the second round of PCR, samples with different barcodes and indexes were mixed, purified by gel extraction using the QIAquick Gel Extraction Kit (Qiagen), and quantified using the Qubit ssDNA HS Assay Kit (Thermo Fisher Scientific). Deep sequencing was performed on the Illumina NovaSeq 6000 platform. Quality control was performed for the sequencing data by fastp (v0.23.2) using default parameters. The sequencing reads were demultiplexed using fastq-multx (v1.4.1) with the barcoded PCR primers, and the editing frequencies of the on-target site were calculated by output file from batch analysis with CRISPResso2 (v2.0.32), and statistics were generated using in-house scripts with R (v4.2.1).

The whole mtDNA was amplified as two overlapping 8 kb fragments by long-range PCR, and the sequence information of the primers was shown in Supplementary Table S2. The PCR products were purified by QIAquick Gel Extraction Kit (Qiagen) and used as input for constructing libraries using TruePrepTM DNA Library Prep Kit V2 for Illumina (Vazyme). The libraries were purified using DNA clean beads and quantified using the Qubit ssDNA HS Assay Kit (Thermo Fisher Scientific) before performing the deep sequencing. To analyze NGS data from whole mitochondrial genome sequencing, the qualified reads were mapped to the mouse mitochondrial reference genome (mm10) by BWA (v0.7.12) with mem–M, and then generated BAM files with SAMtools (v.1.9). Positions with conversion rates ≥0.1% were identified among all cytosines and guanines in the mitochondrial genome using the REDItoolDenovo.py script from REDItools (v.1.2.1).

Single-nucleotide variants present in both treated and untreated samples (that therefore did not arise from DdCBE treatment) were excluded. The average off-target editing frequency was then calculated independently for each biological replicate of each treatment condition as: (number of reads in which a given C•G base pair was called as a T•A base pair, summed over all non-target C•G base pairs)/(total number of reads that covered all non-target C•G base pair).

Figures were drawn with GraphPad Prism 9, Figdraw (www.figdraw.com), and Adobe Illustration 2021. Mean and standard error of measurement (SEM) of editing efficiency of three biological duplicate samples was calculated using SPSS software (version 23.0; SPSS Inc., Chicago, IL, USA). Independent sample t-test was applied to comparison of normally distributed data (ns p > 0.05, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001).

To explore the editing window of mtDNA targets mediated by DdCBEs, nearly one hundred mtDNA target sites were selected, and the location of each site in the mitochondrial genome was marked in Figure 1A (for aC targets) and Figure 1B (for tC targets). To minimize the variation among different targets, a commonly used spacer length with 16 base pairs was designed for all of the selected sites. For aC editing, a total of 61 sites was designed ensuring that there are at least 3 different aC sites for each position between 5′C_3-14 in the spacing region (Figure 1C left panel), and the detailed information of each target is shown in Supplementary Table S3. For each spacer, four DdCBE pairs (G1333CL+G1333NR, G1397CL+G1397NR, G1333NL+G1333CR, and G1397NL+G1397CR) were constructed. After four pairs were transfected into the Neuro-2a cells, the efficiency of C>T conversion at each site was evaluated by the next-generation sequencing (NGS) method, and the results indicated that there is a wide range of mtDNA editing efficiencies (from 1.03% ± 0.46% to 48.99% ± 2.05%) depending on the position of the target C within the spacing region (Supplementary Table S4). When the data for the tC targets were also evaluated with a similar strategy (Supplementary Table S5), it was observed that the editable range of aC targets was more concentrated than that of tC targets, either for the G1333 splits of DddA or for the G1397 splits of DddA. The higher editing efficiency of G1333 pairs or G1397 pairs of each aC target was shown in Figure 1C and that of each tC target was shown in Figure 1D, and the specific editor combination used for each site is listed in Supplementary Table S6. The mean efficiency of C>T conversion for each target cytosine at different positions for aC and tC targets is separately shown in Figures 1E–H. Editing windows are defined when there exist targets with editing efficiency >5% in the position. As Figure 1C indicated, the editing window of aC targets could be defined as 5′C_8-11 for the G1333 split of DddA, while high efficiency (>15%) only could be achieved when the cytosine was put at the position of 5′C_8-10 in the spacer (Figure 1C middle panel; Figure 1E). If aC targets were edited by the G1397 split of DddA, the editing window was defined as 5′C_10-13, but the mean editing efficiency was relatively low (4.39% ± 2.56% to 6.75% ± 3.46%) when the cytosine was put at 5′C_11-13 in the spacer (Figure 1C right panel; Figure 1F). Whereas, when the editing window of the tC targets was evaluated, it was observed that tC targets have a broader editing window than aC targets. As the results showed, tC targets could be efficiently edited when the cytosine was put at 5′C_4-13 in the spacer for the G1333 split of DddA (Figure 1D middle panel), while average editing efficiency (>15%) only could be achieved in 5′C_6-12 (Figure 1G). Meanwhile, the editing window for the tC target was between 5′C_5-14 when it was edited by the G1397 split of DddA (Figure 1D right panel), while the average editing efficiency of more than 15% only could be obtained in 5′C_9-14 (Figure 1H). In summary, all of the results above indicated that the editing window of aC is more concentrated in comparison with that of tC targets, which means aC editing has stricter requirements for spacer selection.

It was found the editing effciency between different orientations of the same loci (G1333CL+G1333NR vs. G1333NL+G1333CR, and G1397CL+G1397NR vs. G1397NL+G1397CR) can have a huge difference. Thus it was suspected whether there exited DddA-half-specificity or DddA-half-preference for editing. DddA-half-specificity was defined as the pair can only edit targets on the specific strand, while the pair of the other orientation can not or barely not edit targets on the strand. DddA-half-preference was defined when pairs derived from the same split can both edit effectively for a specific strand but the editing efficiency of one pair is always higher than that of the other pair with different orientation.

To compare the effect of orientations on the editing efficiency for specific splits of DddA separately to edit aC and tC targets, we used the difference value of editing efficiency between G1333NL+G1333CR pair and G1333CL+G1333NR pair as the vertical axis for G1333 split of DddA, while the difference value of editing efficiency between G1397CL+G1397NR pair and G1397NL+G1397CR pair as the vertical axis for G1397 split of DddA. As Figure 2A indicated, if this value is positive, it means that the editing efficiency of G1333NL+G1333CR pair is higher than G1333CL+G1333NR pair, while if this value is negative, it means that the editing efficiency of G1333NL+G1333CR pair is lower than G1333CL+G1333NR pair. By using this method, it was observed that when G1333 splits of DddA were used for aC editing, the cytosine of aC motif in the top strand can only be edited by G1333NL+G1333CR pair, whereas it almost cannot be edited when the G1333CL+G1333NR pair was used (Figures 2A, B). Remarkably, the case was the opposite when aC motif was in the bottom strand (Figures 2A, C). Herein, it was speculated that G1333 split of DddA might edit aC targets in a half-specific manner. To verify this hypothesis, three additional targets (MT-CO3 site 1; MT-ND3 site 2; MT-ND3 site 3) were selected, and each target has two aC motifs separately located at the top and bottom strands in the same position (the 10th nucleotide) from the 5′ ends of each strand (Figure 2D). When the editing efficiency was evaluated in three different sites, it was also observed that the editing only occurred in the aC motif on the top strand when using G1333NL+G1333CR pair, whereas editing only occurred in the aC motif on the bottom strand when switching the combination (G1333CL+G1333NR pair). Remarkably, these results were perfectly reproducible both in MT-CO3 site 1 (Figure 2E left panel), MT-ND3 site 2 (Figure 2E middle panel) and MT-ND3 site 3 (Figure 2E right panel).

Different from G1333 splits of DddA, both orientations of G1397 splits can edit the aC sites (Figure 3A). To further explore whether the editing scope of different orientations had a preference, we compared the difference value between G1397CL+G1397NR pair and G1397NL+G1397CR pair at different loci (Figures 3B, C). It was found that both orientations could work, but G1397CL+G1397NR pair had significantly higher editing efficiency when the aC motif was on the top strand for most of the selected targets. Oppositely, most of the aC motif on the bottom strand could be edited more effectively by the G1397NL+G1397CR pair. Notably, as Figure 3A indicated, the editable cytosine was mainly concentrated on the right half of the spacing region.

For tC targets, the half-preference only occurred for G1333 but not G1397 splits of DddA (Figures 3D, G). As the results indicated, most of the G1333NL+G1333CR pair had higher editing efficiency when the tC motif in the top strand was edited, while most of the G1333CL+G1333NR pair worked better when the tC motif in the bottom strand was edited (Figures 3E, F). Remarkably, there is no obvious strand bias exists for the G1397 splits of DddA (Figure 3G). No matter which strands the tC motif is located in, both orientations of the G1397 splits of DddA pairs have certain levels of editing efficiency (Figures 3H, I). Interestingly, when calculating the difference value between G1397CL+G1397NR pair and G1397NL+G1397CR pair for all loci (Figure 3G), it was found that tC on the right half of the spacer was edited with priority by the G1397CL+G1397NR pair; while the G1397NL+1397CR pair preferentially edited the tC on the left half of the spacer. Taken together, G1333 split of DddA mediated aC and tC editing separately in half-specific and half-prefer manner, while half-prefer only exists for G1397 split of DddA when it was used for editing aC targets.

According to the results from editing windows screening, it was observed that some sites are difficult to edit even if they are located within the editing window. Thus it was speculated that the nucleotide adjacent to the 3′ end of aC motif might impact the mtDNA editing results. To further confirm our hypothesis, all the aCn motifs in the editing window were classified based on the nucleotide adjacent to the 3′ end (n for t, c, a, or g), and screened mtDNA editing with the different combinations of DdCBE pairs. For each site, all the four DdCBE pairs were constructed and tested. And the editing efficiency of the pair having better editing results of the same split method was used in Figure 4. It was observed that DdCBEs catalyzed the C>T conversion in a motif-dependent manner, and the motif with aCt and aCc can be edited throughout the entire editing windows (5′C_8-11 for the G1333 split and 5′C_10-13 for the G1397 split) (Figure 4A). For the motif with aCa, effective editing (>15%) can only be achieved when placing the target at 5′C_8-10 using the G1333 splits (Figure 4A). As the results summarized in Figures 4B, C, aCt and aCc can achieve relatively higher editing efficiency than aCa and aCg in different positions regardless of the editor combinations, aCa motif could be edited effectively with G1333 splits of DddA only when the cytosine was put at the 5′C_8-10, and the aCg motif could only be edited when the cytosine was put at the 5′C₁₀ in the spacing region either for G1333 or for G1397 splits.

When further exploring the influence of the 3′ end nucleotide on the tC editing, it was found there is no significant association between the specific motif and editing efficiency. According to the currently limited data, the base propensity was not applicable for tC editing, because tC motif with different adjacent nucleotides at the 3′ end (tCa/tCt/tCc/tCg) all work comparably (Figure 4D) for both G1333 and G1397 pairs.

It is known that, due to the restrictions of TALEN design, the length of the spacing region is variable rather than fixed when performing the mtDNA editing in practice. To further investigate the editing rules of the aC targets when the spacer length is more or less than 16 base pairs, some additional aC targets with 18bp (Figure 5A left panel) and 14bp (Figure 5B left panel) in the spacing region were selected for mtDNA editing, and the detailed information of these targets was shown in Supplementary Table S3). For the 18bp spacer, it was found the editing window of G1333 pairs was wider as the spacer length increased – 5′C_7-13, compared with 5′C_8-11 of the 16bp spacer (Figure 5A middle panel; Figure 5C). The editing window of G1397 pairs was between 5′C_11-13 for the 18bp spacer compared with 5′C_10-13 for the 16bp spacer (Figure 5A right panel; Figure 5D). Correspondingly, for the 14bp spacer, the editing window of G1333 pairs was relatively shorter -5′C_7-10 (Figure 5B middle panel; Figure 5E), while aC can be edited by G1397 splits of DddA only when the cytosine was put at 5′C₉ and 5′C₁₀ (Figure 5B right panel; Figure 5F). Remarkably, there were also exists some uneditable sites within the editing window both for the 18bp and the 14bp spacer, but most of which are aCa and aCg sites (such as sites 18S-6, 18S-7, 18S-8, 18S-9, 18S-10, 18S-15, and 18S-25 in Figure 5A and 14S-5, 14S-9, 14S-10, and 14S-11 in Figure 5B), which suggested that the effect of nucleotide adjacent to the 3′ end of aC motif also exists for spacers with 18 and 14 base pairs. Meanwhile, it was observed that the DddA-half-dependent editing specificity of G1333 pairs applied equally to the 18bp and the 14bp spacers (Supplementary Figure S1). However, the editing preference of G1397 pairs only could be observed in 18bp spacer, but not in 14bp spacer (Supplementary Figure S1). The detailed editing efficiency of each site is shown in Supplementary Table S7. Taken together, the length of the spacing region only affects the position of editable cytosine for aC targets, while the effect of half preference and nucleotide adjacent to the 3′ end of aC motif also existed for spacers of different length.

It is known that, off-target effects are another factor that affects the application of DdCBEs editors. If there exists more than one option for TALEN binding sequence for mtDNA editing of certain target, it still uncertain how to choose to improve editing specificity. In 2017, Rinaldi and his colleagues used a series of synthetic TALEs to investigate how the number of TALEN RVD repeats affects the its binding affinity to the target and non-target, and the results demonstrated that the specificity (the ratio of affinity for target DNA to affinity for non-target DNA) was variable with increasing number of RVDs, and the optimal length for specificity was between 15 and 19 RVDs (Rinaldi et al., 2017). In this work, to clarify whether the number of RVDs affects the off-target during DdCBE mediated mtDNA editing, two target sites were selected (mt.9545, mt.7155). For each site, three different pairs of TALEN binding sequences (Strategy-1, Strategy-2, and Strategy-3) were named based on the length of the TALEN binding sequence, from short to long, to compare the off-target effects (Figure 6A). To control variables, same split and orientation was used in different strategy (G1333CL+G1333NR for mt.9545, and G1333NL+G1333CR for mt.7155, which was the pair with the highest editing efficiency for each site respectively). As the results indicated, Strategy-1 (TALEN binding sequences of 16bp and 14bp) led to lower average off-target editing frequency than Strategy-2 and 3 (Figure 6B) when three strategies were used for mtDNA editing at mt.9545 site. Similarly, for mt.7155, Strategy-1 (TALEN binding sequences of 15bp and 14bp) led to the lowest average off-target editing frequency (Figure 6B). The detailed information of mitochondrial genome-wide off-targets for this two sites was separately shown in Figures 6C, D. The sequence information of off-target sites of three strategies of mt.9545 and mt.7155 was shown in Supplementary Figure S2. All the off-target sequences (>1%) identified in each strategy were put together to find test sequence similarity. And it was found that there was no similarity between the upstream and downstream sequences of off-target sites and the TALEN recognition sequences of on-target sites. Based on our limited data, it is suspected the number of RVDs may influence mtDNA editing specificity, and sixteen RVDs and above might increase the risk of mitochondrial genome-wide off-targets. This may due to the occurrence of extra binding force caused by excessive RVDs, which might lead to more severe off targets on the entire mitochondrial genome (Rinaldi et al., 2017). However, owing to the limited data, further experiments are needed to verify this suspection.

To explore whether the different DdCBE pair have impact on off-target editing, three target spacers were chosen (target-1, target-2, target-3), and the mitochondrialgenome-wide off-targets among the four pairs of the same spacer were compared. For all the three spacers, G1397 pairs led to lower average off-target editing frequency than G1333 pairs (Figure 6E). And the number of off-targets was also much more in cells edited by G1333 pairs than in cells edited by G1397 pairs (Figure 6F). Thus it is suspected that G1397 pairs have higher editing specificity.

For the convenience for clinical application in the future, we efficiently constructed a cell model containing an identified pathogenic mtDNA mutation located on the MT-TH gene (mt.12147 G>A) under the guidance of the editing rules above. As the mtDNA sequence shown in Figure 7A, besides the target aC motif (marked in red) on the bottom strand, there still exists two additional editable aC sites (marked in green) separately located in the top and bottom strand within the spacing region, which might cause the bystander effects. According to the traditional procedure, a great many time- and effort-consuming screening works need to be performed, including designing different kinds of target spacers and the corresponding four combinations of DddA for each potential target spacer. However, according to our findings above, aC motif on the bottom strand can be effectively edited by the G1333CL+G1333NR pair when the cytosine was put at 5′C_8-10 in the 16bp spacer. Consequently, we placed the target cytosine at the 5′C₉ within the 16bp-spacing region (Strategy A). All of the four DdCBE pairs were constructed and editing efficiency was compared (Figure 7B). As expected, only the G1333CL+G1333NR pair could mediate C>T conversion effectively at the mt.12147 site in human U2-OS cells. On the other hand, there was no bystander effect at the potentially editable site (5′C₁₄) within the spacer, since this non-target cytosine was out of the editing window 5′C_8-11 for G1333 splits of DddA. However, another two pairs of TALE arrays were engineered, for which the target cytosine was placed at position at 5′C₄ and the 5′C₁₃ within the spacing region respectively (Strategy B and C), and it was observed that mt.12147 has hardly been edited in four pairs when the other two strategies were used in this work (Figures 7C, D). Remarkably, for strategy C, bystander effects occurred as the non-target cytosine was placed at the position of 5′C₁₀, which is within the editing window for both G1333 and G1397 splits of DddA. In summary, the optimal strategy conceived based on our findings did achieve the best editing efficiency without the bystander effects.