Nuclear and mitochondrial DNA editing in human cells with zinc finger deaminases

Base editing in nuclear DNA and mitochondrial DNA (mtDNA) is broadly useful for biomedical research, medicine, and biotechnology. Here, we present a base editing platform, termed zinc finger deaminases (ZFDs), composed of custom-designed zinc-finger DNA-binding proteins, the split interbacterial toxin deaminase DddAtox, and a uracil glycosylase inhibitor (UGI), which catalyze targeted C-to-T base conversions without inducing unwanted small insertions and deletions (indels) in human cells. We assemble plasmids encoding ZFDs using publicly available zinc finger resources to achieve base editing at frequencies of up to 60% in nuclear DNA and 30% in mtDNA. Because ZFDs, unlike CRISPR-derived base editors, do not cleave DNA to yield single- or double-strand breaks, no unwanted indels caused by error-prone non-homologous end joining are produced at target sites. Furthermore, recombinant ZFD proteins, expressed in and purified from E. coli, penetrate cultured human cells spontaneously to induce targeted base conversions, demonstrating the proof-of-principle of gene-free gene therapy. Base editing in nuclear DNA and mitochondrial DNA (mtDNA) is broadly useful for biomedical research, medicine, and biotechnology. Here the authors present zinc finger deaminases which catalyze targeted C-to-T base conversions without inducing unwanted indels in human cells.

A growing list of tools have been reported for genome editing in eukaryotic cells and organisms, a broadly useful method in research and medicine, which now include, but are not limited to, zinc finger nucleases (ZFNs) 1 , transcription activator-like effector (TALE) nucleases (TALENs) 2 , TALElinked split interbacterial deaminase toxin DddA-derived cytosine base editors (a.k.a. DdCBEs) 3 , CRISPR-Cas9 4-6 , and catalytically-impaired Cas9-linked deaminases (a.k.a. base editors) [7][8][9] . These tools, in principle, are composed of two functional units: a DNA-binding moiety and a catalytic moiety. Thus, a zinc finger array or a TALE array functions as a DNA-binding moiety, whereas a nuclease (FokI in ZFNs and TALENs) or a deaminase enzyme (split-DddA tox in DdCBEs and APOBEC1 in CBEs) functions as a catalytic unit. CRISPR-Cas9 is both a nuclease enzyme and an RNA-guided DNA-binding protein.
Custom-designed, programmable nucleases such as ZFNs, TALENs, and Cas9 cleave DNA, producing double-strand breaks (DSBs), the repair of which give rise to gene knock-out and knock-in in a targeted manner. Programmable nuclease-induced DSBs, however, can cause unwanted large deletions 10-13 at ontarget sites, p53 activation [14][15][16] , and chromosomal rearrangements 17,18 resulting from the repair of two concurrent DSBs at on-and off-target sites. In contrast, programmable deaminases, including cytosine and adenine base editors (CBEs 7,9 and ABEs 8 ), do not produce DSBs, avoiding these unwanted events in cells, and catalyze single nucleotide conversions efficiently without a repair template or donor DNA. Note, however, that CBEs and ABEs, containing the Cas9 nickase variant, cleave the target DNA strand and produce nicks or single-strand breaks, which can still cause unwanted indels at target sites.
Recently, Mok et al. 3 demonstrated that the interbacterial deaminase toxin DddA tox derived from Burkholderia cenocepacia can be split and fused to TALE arrays and a uracil glycosylase inhibitor (UGI) to create DdCBEs, which catalyze C-to-T base conversions in nuclear DNA and mitochondrial DNA (mtDNA) in mammalian cells. We and others also showed mtDNA editing in mice 19 and chloroplast DNA editing in plants 20,21 using custom-designed DdCBEs. In this study, we sought to create zinc finger deaminases (ZFDs) for indel-free, precision base editing in human and other eukaryotic cells by linking split-DddA tox to custom-designed zinc finger proteins (ZFPs). Because zinc finger arrays (encoded in a 2 × 0.3~0.6 k base pair (kbp) DNA) are small in size, compared to TALE arrays (2 × 1.7~2 kbp) or S. pyogenes Cas9 (4.1 kbp) 22 , ZFD-encoding genes can readily be packaged in a viral vector with a limited cargo space such as AAV for in vivo studies and gene therapy applications. Unlike TALE arrays, zinc finger arrays lack bulky domains at both the C terminus and the N terminus, making them engineering friendly: split-DddA tox halves can be fused to either terminus of a ZFP. Furthermore, ZFPs with an intrinsic cell-penetrating activity [23][24][25][26] may allow nucleic acid-free gene editing in human cells. These properties may make ZFPs an ideal platform as a DNA-binding module for base editing in nuclear and organelle DNA.

Results
Optimization of ZFD architecture. To develop ZFDs for base editing in human and other eukaryotic cells, we first sought to optimize both the length of the amino-acid (AA) linkers that connect ZFPs to split-DddA tox halves and the length of the spacers, where C-to-T conversions would be induced, between the left and right ZFP-binding sites. We chose a well-characterized ZFN pair, targeted to the human CCR5 gene, to make ZFDs with variable linkers of 2, 5, 10, 16, 24, and 32 AA residues in length and constructed a series of target plasmids (pTargets) that contained left and right ZFP-binding sites separated by variable spacers, composed of repetitive 'TC' motifs, of one to 24 base pairs (bps) in length (Fig. 1a, b and Supplementary Table 1,2). Note that DddA tox can be split at two positions (G1333 and G1397) and that each half can be fused to either a left or right ZFP. We measured the base editing efficiencies of the resulting 24 (= 6 linkers × 2 split positions × 2 possible fusions (left or right)) ZFD constructs with each of the 24 pTarget plasmids in HEK 293 T cells at day four post-transfection via targeted deep sequencing. ZFD constructs with short linkers (2-and 5-AA in length) were poorly efficient or inactive, whereas those with linkers of at least 10-AA in length induced C-to-T conversions with pTarget plasmids containing spacers of at least 4-bp in length at frequencies that ranged from 1-24% ( Fig. 1c and Supplementary  Fig. 1a, b). ZFD pairs with a 24-AA linker showed the highest editing efficiencies. To determine the best combinations of linkers, we also measured the editing efficiencies of ZFD pairs that consisted of a left ZFD construct with the 24-AA linker and a right ZFD with variable linkers or vice versa ( Fig. 1d and Supplementary Fig. 2). We found that the right ZFD with the 24-AA linker was most active when paired with the left ZFD with the same 24-AA linker. We also found that ZFDs with DddA tox split at G1397 were more efficient than those with DddA tox split at G1333 ( Fig. 1c and Supplementary Fig. 1a, b). Cytosines were edited by these most efficient ZFD pairs with high efficiencies of >6.8% in spacer regions of 7-21 bps in length ( Fig. 1c and Supplementary Fig. 1a, c).
Base editing at endogenous target sites. Next, we investigated whether ZFDs with the 24-AA linker could catalyze C-to-T edits at endogenous chromosomal target sites in human cells. A total of 22 ZFD pairs were designed to target 11 sites (two ZFD pairs per site) in eight genes (Fig. 2 and Supplementary Table 3, 4). Among them, 14 ZFD pairs were constructed from scratch using publically-available zinc finger resources. We also modified previously-characterized ZFNs (specific to CCR5 27 and TRAC 28 ) to make the other eight ZFD pairs. Because ZFN pairs cleave target DNA in spacer regions of 5-7 bps 22 in length, whereas our ZFD pairs preferentially operate in spacer regions of at least 7 bps, we deleted one or two zinc fingers from these ZFN pairs and added a few zinc fingers to make ZFDs ( Supplementary Fig. 3). Since the FokI nuclease domain can be fused to either the N-or C-termini of ZFPs to create ZFNs with four different configurations 28 , we also constructed two ZFD pairs (TRAC-NC in Fig. 2b) with an alternative configuration (shown as NC configuration in Fig. 2a and Supplementary Fig. 4) to test whether split-DddA tox halves can be fused to the N terminus of ZFPs.
In HEK 293 T cells, efficiencies of C-to-T base editing by these ZFDs, including those with the NC configuration, ranged from 1.0% to 60%, whereas indels were rarely induced, showing frequencies of <0.4% ( Fig. 2b and Supplementary Fig. 5). As anticipated from our plasmid-based assays described above, two ZFD pairs targeted to the CCR5 site with a 5-bp spacer were poorly efficient. The other 20 ZFD pairs targeted to sites with spacers of at least 7 bps showed an average editing frequency of 12.0 ± 3.4%, on par with Cas9-derived Base Editor 2 (8.3 ± 2.2%) 29 . In addition to cytosines in the context of TC, those in the context of AC and GCC were also converted to thymine, albeit less efficiently (Fig. 2c-f). Thus, C 6 in the context of AC at the NUMBL site and C 7 in the context of GCC at the INPP5D-2 site were converted to T with frequencies of up to 4.6 and 1.9%, respectively.
Direct delivery of purified ZFD proteins into human cells. Delivery of purified recombinant gene-editing enzymes, rather than plasmid DNA encoding them, into cells can reduce offtarget effects, avoid innate immune responses against foreign ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-022-27962-0 DNA, and exclude the possibility of integration of plasmid DNA fragments into the genome. We and others have shown that ZFPs can spontaneously penetrate into mammalian cells both in vitro 24,25 and in vivo 23 . To demonstrate ZFD protein-mediated base editing in cultured human cells, we chose the highly active ZFD pair targeted to the TRAC site (TRAC-NC) and purified recombinant TRAC-NC proteins, containing one or four copies of a nuclear localization signal (NLS), from E. coli. We first assessed the deaminase activity of the ZFD proteins in vitro using a PCR amplicon containing the TRAC site and found that they were highly active, showing efficient DNA cleavage upon treatment with Uracil-Specific Excision Reagent (USER), a mixture of Uracil DNA glycosylase and DNA glycosylase-lyase Endonuclease VIII ( Supplementary Fig. 6). We then delivered the TRAC-NC ZFD proteins into human leukemia K562 cells via two different methods, electroporation or direct delivery without electroporation. These ZFD proteins were highly efficient, inducing targeted C-to-T conversions at frequencies of up to 27% (electroporation)

T G G T C A T C C T C A T C … … T C … … G A … … A A A C T G C A A A A G G C A C C A G T A G G A G T A G … … A G … … C T … … T T T G A C G T T T T C C G
Mitochondrial DNA editing with ZFDs. One major advantage of the split-DddA tox system fused to custom-designed DNAbinding proteins is that these programmable deaminases, unlike CRISPR-based systems, can be used for editing organelle DNA

C G A A T C C T C C T G C T T A G G A G G A
Left Right including mtDNA 3 and chloroplast DNA 20,21 . To deliver ZFDs to mitochondria, we constructed mitoZFDs (mitochondria-targeting ZFDs) by linking the mitochondrial targeting signal (MTS) and nuclear export signal (NES) sequences 30 to the N terminus of custom-designed ZFDs targeted to eight mitochondrial genes ( Supplementary Fig. 7). ZFP-encoding DNA segments (full sequences are shown in Supplementary Table 5, 6) for these ZFDs were assembled using a publicly available zinc finger resource. These ZFDs were designed to recognize left-and right-binding sites of 12 bps in length separated by a spacer of 7-15 bps. The efficiencies of mtDNA editing in HEK 293 T cells by these mitoZFDs ranged from 2.6-30% (11 ± 2% on average, n = 18) (Fig. 3a). Interestingly, mitoZFDs with the NC configuration (NC ZFDs) were more efficient (13 ± 3%, n = 12) than those with the CC configuration (CC ZFDs) (7 ± 2%, n = 6). (However, this does not mean that NC ZFDs, in general, are advantageous over CC ZFDs, because different ZFPs are used in NC ZFDs and CC ZFDs. It is possible that CC ZFDs but not NC ZFDs (or vice versa) can be designed to recognize a pre-determined target site.) Most cytosines in the TC context (and those in the TCC context, which can be converted to TTC first and then to TTT) in a spacer region were converted, with variable efficiencies (Fig. 3b-g). In addition, two cytosines (C 8 and C 9 ) in the ND2 site in the context of ACC were edited with fair frequencies of up to 12 and 11%, respectively (Fig. 3b), which suggests that ZFD-mediated C-to-T editing is not limited to the TC motif. We next isolated single cell-derived clonal populations of mtDNA mutant cells to demonstrate that mitoZFDs are not cytotoxic and that mtDNA mutations are stably maintained after clonal expansion. Among 30 single cell-derived clones obtained from HEK 293 T cells treated with the ND1-specific mitoZFD, five clones contained mutations in the ND1 gene with frequencies ranging from 35-98% ( Supplementary Fig. 8a). Likewise, among 36 single cell-derived clones isolated from ND2-mitoZFD-treated cells, we obtained seven base-edited clones that contained ND2 mutations with frequencies that ranged from 26-76% (Supplementary Fig. 8b). All of the other clones had mutations at low frequencies of 0.4-1.0%, most likely resulting from sequencing errors: ZFD-untreated cells had mutations at similar frequencies ( Supplementary Fig. 8c). These results show that mitoZFDs induced heteroplasmic mutations unevenly in a population of cells: Thus, most of the ZFD-transfected cells were

T A C T C C A C C T C A A A T G A G G T G G A G T T
Left Right    mitoZFDs and TALE-based DdCBEs. We found that mutation patterns induced by our ND1-specific mitoZFDs were quite different from those induced by two TALE-based DdCBE pairs designed to target the same site (Fig. 3f-h). Thus, the two mitoZFDs catalyzed C-to-T conversions at the C 5 or C 8 position (Fig. 3f), whereas the DdCBEs induced base conversions at C 8 , C 9 , and C 11 positions (Fig. 3g). As a result, amino-acid changes caused by mitoZFDs were completely different from those caused by DdCBEs (Fig. 3h). Note that our mitoZFDs bind to left-and right-half sites separated by an 8-bp spacer, whereas DdCBEs bind to target sites separated by a 16-bp spacer, which is likely responsible for the differential mutation patterns. These results suggest that mitoZFDs and DdCBEs can be used to generate complementary, diverse mutation patterns in mtDNA.

A C T C A A T C C T C T G A T C T G A G T T A G G A G A C T A G
To further expand possible mutation patterns, we tested whether a ZFD monomer could be combined with a DdCBE monomer to create a functional hybrid pair (Fig. 4a). Ten ZFD/ DdCBE hybrid pairs specific to the ND1 site were highly efficient in HEK 293 T cells, with an average editing frequency of 17 ± 3% (Fig. 4b). In fact, one of the hybrid pairs (TALE-L/ZFD-R1) was more efficient than the two DdCBE pairs and ten ZFD pairs targeted to the same site and exhibited the highest editing frequency (41%) (Fig. 4b). Furthermore, hybrid pairs yielded mutation patterns different from those obtained with ZFDs or DdCBEs (Fig. 4c). Of note, a few hybrid pairs (for example, ZFD-L1/TALE-R and ZFD-L2/TALE-R) induced C-to-T edits at single positions without bystander edits. In contrast, most of the ZFD pairs and DdCBE pairs induced C-to-T conversions at multiple positions in spacer regions. We also found that ZFD/DdCBE hybrid pairs designed to target a Cox2 site were as efficient as ZFD pairs and DdCBE pairs targeted to the same site and produced mutation patterns distinct from those obtained with ZFD pairs and DdCBE pairs ( Supplementary Fig. 11). These results demonstrate that ZFD/DdCBE hybrid pairs can create unique mutation patterns and produce certain mutations that cannot be obtained using ZFDs or DdCBEs.
Mitochondrial genome-wide target specificity of ZFDs. To investigate whether mitoZFDs exhibit off-target DNA editing in mtDNA, we performed whole mitochondrial genome sequencing using DNA samples isolated from cells transfected with ND1-or ND2-targeted mitoZFD pairs. Variable amounts (5-500 ng) of mRNA or plasmids encoding these ZFD pairs were transfected into HEK 293 T cells. As expected, on-target editing frequencies were dose-dependent. High doses (100, 200, and 500 ng) of mRNA or plasmids yielded on-target C-to-T edits with frequencies of >30% but also caused hundreds of off-target edits with frequencies of >1.0% in the mitochondrial genome (Supplementary Figs. 12-15). Sequence logos obtained at off-target sites of the ND2-specific mitoZFD showed a preference for the TC context, indicative of the DddA tox substrate specificity (Supplementary Fig. 16). Low doses (5 and 10 ng) of mRNA or plasmids largely avoided these off-target edits but reduced ontarget mutation frequencies significantly. A medium dose (50 ng) of mRNA was optimal, avoiding hundreds of off-target edits while maintaining high on-target mutation frequencies. To further eliminate remaining off-target edits, we incorporated R(-5)Q mutations in each zinc finger in the ZFDs to remove non-specific DNA contacts 31 . The resulting ZFD variant pair (shown as QQ in Fig. 5) retained high on-target activity and showed exquisite specificity with almost no off-target edits, compared to untreated mtDNA. In particular, the specificity ratio was improved by 8.2fold, when 50 ng mRNA of mitoZFD with R(-5)Q mutations was used, compared to 200 ng of WT mitoZFD mRNA (Fig. 5e). Additionally, we assessed off-target editing in nuclear DNA at sites with high sequence homology to ZFP-binding sequences. No off-target edits were detectably induced by the ND4L mitoZFD at three homologous sites that differ from the on-target site by a single or two nucleotides ( Supplementary Fig. 17a), whereas offtarget edits were induced with low (~1.0%) frequencies by the ND2 mitoZFD at a homologous site with a one-nucleotide mismatch ( Supplementary Fig. 17b). Use of the QQ variant pair reduced off-target edit frequency to 0.4% at this site (Supplementary Fig. 17b).

Discussion
Base editing is a relatively new method for generating targeted nucleotide substitutions without double-strand DNA cleavage or a repair DNA template. Base editing enables C-to-T or A-to-G conversions in cell lines 3,7,8 , animals 19,32,33 , and plants 20,21,34 , allowing researchers to study the functional effects of singlenucleotide polymorphisms (SNPs) and to correct disease-causing point mutations for therapeutic applications. Two types of base editors have been developed: CRISPR-derived ABEs and CBEs and DddA-derived CBEs. CRISPR-derived base editors are composed of catalytically-impaired Cas9 or Cas12a variants as DNA-binding units and single-strand DNA-specific deaminases originated from rat, sea lamprey, or E. coli 7-9,35 , whereas DdCBEs are composed of TALE DNA-binding arrays and double-strand DNA-specific DddA tox 3 . In this study, we presented an alternative type of base editors, ZFDs, composed of zinc finger DNA-binding arrays and DddA tox .
Compared to DdCBEs, ZFDs are smaller in size, because the ZFPs in ZFDs are compact, whereas the TALE arrays in DdCBEs are bulky. As a result, a ZFD pair-encoding gene but not a DdCBE pair-encoding gene can be readily packaged in an AAV vector with a small cargo space. In addition, compact ZFPs are engineering friendly, making it possible to fuse split-DddA tox halves to either the C or N terminus of a ZFP, creating ZFDs that operate either upstream or downstream of a ZFP-binding site. Furthermore, recombinant ZFD proteins can penetrate into human cells spontaneously without electroporation or lipofection, potentially enabling gene-free gene therapy. Last but not least, ZFD pairs or ZFD/DdCBE hybrid pairs can produce unique mutation patterns, which cannot be obtained using DdCBEs alone. These properties will make ZFDs a powerful platform for modeling and treating mitochondrial diseases.
We expect that ZFDs can be further engineered to improve efficiency and specificity. Here, we showed that use of ZFP variants and ZFD mRNA can reduce off-target activity. DddA tox can also be engineered to avoid ZFD off-target mutations. ZFDs with enhanced efficiency and precision could pave the way for correcting pathogenic mitochondrial DNA mutations in human embryos, fetuses, and patients.
The pTarget plasmids were designed for determining the optimal length of the spacer sequence for ZFD activity. Each pTarget plasmid, which contains two ZFPbinding sites with a spacer of variable length between them, was constructed by inserting the ZFP-binding sequences and a spacer sequence into the pRGS-CCR5-NHEJ reporter plasmid after it had been digested with two enzymes (EcoRI and BamHI, NEB), which recognize sites between the RFP and EGFP sequences.
pET-ZFD plasmids for protein production in E. coli were created by modifying the pET-Hisx6-rAPOBEC1-XTEN-nCas9-UGI-NLS plasmid (addgene, #89508) 33 after digestion with NcoI and XhoI (NEB). ZFD sequences were amplified from the p3s-ZFD plasmid using PCR, and Hisx6 tag and GST tag sequences were synthesized as oligonucleotides (Macrogen). All plasmids were generated using a HiFi DNA assembly kit (NEB) to insert sequences encoding the ZFD and tag for protein purification into the digested pET plasmid.
DH5ɑ chemically competent E. coli cells were used for transformation of plasmids, and plasmids were purified with an AccuPrep Plasmid Mini Extraction Kit (Bioneer) according to the manufacturer's protocol. The desired plasmids were selected after confirming the entire sequence with Sanger sequencing.
ZFD design scheme. In this paper, ZFDs were prepared in two ways. First, ZFDs were made by removing or attaching zinc finger protein from previously constructed ZFN. Unlike the ZFN whose optimal spacer is 5~6 bp, the optimal spacer of ZFD is ≥7 bp, so the zinc finger on the back part was removed to widen the spacer. When manufacturing zinc finger array protein, it is recommended to use four or more fingers. If there were <4 after removing one zinc finger, we added one zinc finger to the front part ( Supplementary Fig. 3). Second, ZFDs were de novo  assembled using a publicly available zinc finger resource (from Toolgen 36 ). In this resource, 33 ZFP were recommended for use. In this way, since the binding ability of the ZFP array is considered critical for ZFD activity, choose the N-type or C type that can use the recommended ZFP at the DNA binding site. Finally, the generated ZFP was cloned using different original plasmids depending on the nucleus and mitochondria targets (Supplementary Figs. 4 and 7). For whole-mtDNA sequencing, HEK 293 T cells were transfected with serially diluted concentrations of plasmid or mRNA encoding ND1-or ND2-targeted mitoZFD pairs. Because more cells were required for mtDNA extraction than for analysis of the editing efficiency at defined sites, four samples, transfected under the same conditions as described above, were collected as a single sample (at four times (If no R was present at position-5 of the zinc finger framework, a nearby K or R was converted to Q.) b Whole-mtDNA sequencing of mitoZFD-treated cells. Editing frequencies at on-target and off-target sites are indicated by red and black dots, respectively. Error bars are shown as standard error of the mean (s.e.m.) for n = 2 biologically independent samples. All C/G-to-T/A base changes present at frequencies >1% are presented. c-e Editing efficiencies and specificities depend on the dose of ZFD-encoding mRNA delivered. Data are shown as means from n = 2 biologically independent samples. c The average C/G-to-T/A editing frequency for all C/Gs in the mitochondrial genome. d The number of edited C/G sites with base editing frequencies >1%. e The specificity ratio was calculated by dividing (average editing frequency at on-target Cs) by (average editing frequency at off-target Cs). Source data are provided as a Source Data file. the scale). In the manuscript, the amounts of constructs (ng) that were delivered per 7.5 × 10 4 cells are indicated. mtDNA was isolated from cells at 96 h post transfection.
For ZFD delivery into K562 cells by electroporation, an Amaxa 4D-Nucleofector™ X Unit system with program FF-120 (Lonza) was used. The maximum volume of substrate solution added to each sample was 2 μL when using a 16-well Nucleocuvette™ Strip. 220 pmol (for maximum capacity) or 110 pmol (for half of the maximum capacity) of each of the left and right ZFD proteins, or 500 ng of plasmid encoding left and right ZFD, were transfected into K562 cells (1 × 10 5 ). At 96 h post treatment, cells were collected by centrifugation at 100 g for 5 min, and lysed by incubation at 55°C for 1 h, and then at 95°C for 10 min, in 100 μL of cell lysis buffer (50 mM Tris-HCl, pH 8.0 (Sigma-Aldrich), 1 mM EDTA (Sigma-Aldrich), 0.005% sodium dodecyl sulfate (Sigma-Aldrich)) supplemented with 5 μL of Proteinase K (Qiagen).
For direct delivery of ZFD or ZFD-encoding plasmids into K562 cells, we referred to a method previously used for direct delivery of ZFN 24 . A mixture of left and right ZFD proteins (at a final concentration of 50 μM) or a mixture of plasmids encoding left and right ZFD (500 ng each) was diluted into serum-free medium containing 100 mM L-arginine and 90 μM ZnCl 2 at pH 7.4 to a final volume of 20 μL. K562 cells (1 × 10 5 ) were centrifuged at 100 g for 5 min, and the supernatant was removed. The cells were then resuspended in the diluted ZFD solution and incubated for 1 h at 37°C. After incubation, cells were centrifuged at 100 g for 5 min, and then resuspended in fresh culture medium. Cells were maintained at 30°C (for a transient hypothermic condition) or 37°C for 18 h, and then for two more days at 37°C. Some cells were subjected to a second treatment, following the above process. Cells were analyzed 96 h after treatment.
ZFD protein expression and purification. The plasmids encoding each pair of ZFDs (Supplementary Table 7), each with a C-terminal GST tag, were transformed into Rosetta (DE3) competent cells, which were then cultured on LB-agar plates containing 50 µg/ml kanamycin. After incubation overnight, a single colony was picked and grown overnight (pre-culture) in LB broth containing 50 µg/ml kanamycin and 100 µM ZnCl 2 at 37°C. The next day, part of the pre-culture was transferred to a large volume of LB broth, which was incubated at 37°C with shaking at 200 rpm until the absorbance, A600 nm =~0.5-0.70. The cultures were put on ice for about 1 h, after which ZFD protein expression was induced by the addition of 0.5 mM of isopropyl β-D-1-thiogalactopyranoside (IPTG; GoldBio) and the culture was incubated at 18°C for 14-16 h.
In vitro deamination of PCR amplicons by ZFD. An amplicon containing the TRAC site (Supplementary Table 8) was prepared using PCR. Eight microgram of the amplicon was incubated with 2 µg of each ZFD protein (Left-G1397-N and Right-G1397-C) in NEB3.1 buffer containing 100 µM ZnCl 2 for 1-2 h at 37°C. Following the reaction, ZFD proteins were removed by incubating with 4 µL of Proteinase K solution (Qiagen) for 30 min at 55°C, and the amplicon was purified using a PCR purification kit (MGmed). One microgram of the purified amplicon was incubated with 2 units of USER enzyme (NEB) for 1 h at 37°C. Then, the amplicon was incubated with 4 μL of Proteinase K solution (Qiagen) and purified again using a PCR purification kit (MGmed). The product was subjected to electrophoresis on an agarose gel and imaged. mRNA preparation. DNA templates containing a T7 RNA polymerase promoter upstream of the ZFD sequence were generated from p3s-ZFD plasmids by PCR amplification using Q5 high fidelity DNA polymerase (NEB) with forward and reverse primers (Forward: 5'-CATCAATGGGCGTGGATAG-3', Reverse: 5'-GACACCTACTCAGACAATGC-3'). mRNAs were synthesized in vitro using an mMESSAGE mMACHINE™ T7 ULTRA Transcription Kit (Thermo Fisher). In vitro transcribed mRNAs were purified using a MEGAclear™ Transcription Clean-Up Kit (Thermo Fisher) according to the manufacturer's protocol.
Targeted deep sequencing. Nested PCR was used to produce libraries for nextgeneration sequencing (NGS). The region of interest was first amplified by PCR using KAPA HiFi HotStart PCR polymerase (Roche). To generate NGS libraries, amplicons were amplified again using TruSeq DNA-RNA CD index-containing primers to label each fragment with adapter and index sequences. Final PCR products were purified using a PCR purification kit (MGmed) and sequenced using a MiniSeq sequencer (Illumina) with a GenerateFASTQ workflow. Primer sequences for targeted deep sequencing are listed in Supplementary Table 10. Substitution and indel frequencies from targeted deep sequencing data were calculated with source code (https://github.com/ibs-cge/maund, written by BotBot Inc.).
Whole mitochondrial genome sequencing. For whole mitochondrial genome sequencing, three steps were required: mtDNA extraction from isolated mitochondria, NGS library generation, and NGS. First, 3 × 10 5 HEK 293 T cells were trypsinized and collected by centrifugation (500 g, 4 min, 4°C) 96 h after transfection with ND1-or ND2-targeted mitoZFD pairs. Then, cells were washed with ice-cold phosphate-buffered saline (Welgene), and collected again by centrifugation. The supernatant was removed, and the mitochondria were isolated from cultured cells using the reagent-based method of the Mitochondria Isolation Kit for Cultured Cells (Thermo Fisher) according to the manufacturer's protocol. mtDNA was extracted from isolated mitochondria with a DNeasy Blood & Tissue Kit (Qiagen). To generate an NGS library from the extracted mtDNA, we used an Illumina DNA Prep kit with Nextera™ DNA CD Indexes (Illumina). Finally, the libraries were pooled and loaded onto a MiniSeq sequencer (Illumina). The average sequencing depth was >50.
Analysis of mitochondrial genome-wide DNA editing. To analyze NGS data from whole mitochondrial genome sequencing, we referred to a method previously used for DNA off-target analysis of TALE-DdCBE 3 . First, we aligned the Fastq files to the GRCh38.p13 (release v102) reference genome using BWA (v.0.7.17), and generated BAM files with SAMtools (v.1.9) by fixing read pairing information and flags. Then, we used the REDItoolDenovo.py script from REDItools (v.1.2.1) 38 to identify, among all cytosine and guanines in the mitochondrial genome, the positions with conversion rates ≥1%. We excluded positions with conversion rates ≥50% in all samples, regarding these as single-nucleotide variations in the cell lines. We also excluded the on-target sites for each ZFD treatment. We considered the remaining positions to be off-target sites and counted the number of edited C/G nucleotides with an editing frequency ≥1%. We averaged the conversion rates at each base position in the off-target sites to calculate the average C/G to T/A editing frequency for all C/Gs in the mitochondrial genome. Specificity ratios were calculated by dividing the average on-target editing frequency by the average offtarget editing frequency. Mitochondrial genome-wide graphs were created by plotting the conversion rates at on-target and off-target sites.  Table. Source data are provided with this paper. The plasmid encoding each ZFD pair for TRAC-NC site in nuclear DNA and ND1 and ND2 site in mitochondrial DNA are available from Addgene (TRAC-NC ZFD pair; Addgene #180772 and #180773, ND1 mitoZFD pair; Addgene #180768 and #180769, ND2 mitoZFD pair; Addgene #180770 and #180771). Any other additional relevant data are available from the authors upon request.

Code availability
Base editing frequencies and indel frequencies from targeted deep sequencing data were calculated with source code (https://github.com/ibs-cge/maund, written by BotBot Inc.).