Trends and prospects in mitochondrial genome editing

Mitochondria are of fundamental importance in programmed cell death, cellular metabolism, and intracellular calcium concentration modulation, and inheritable mitochondrial disorders via mitochondrial DNA (mtDNA) mutation cause several diseases in various organs and systems. Nevertheless, mtDNA editing, which plays an essential role in the treatment of mitochondrial disorders, still faces several challenges. Recently, programmable editing tools for mtDNA base editing, such as cytosine base editors derived from DddA (DdCBEs), transcription activator-like effector (TALE)-linked deaminase (TALED), and zinc finger deaminase (ZFD), have emerged with considerable potential for correcting pathogenic mtDNA variants. In this review, we depict recent advances in the field, including structural biology and repair mechanisms, and discuss the prospects of using base editing tools on mtDNA to broaden insight into their medical applicability for treating mitochondrial diseases.


INTRODUCTION
Mitochondria, double membrane-bound organelles referred to as the "powerhouses of the cell", play an indispensable role in eukaryotic cells, as they are associated with metabolism 1,2 and orchestrate a variety of other cellular functions, such as apoptosis, cell pluripotency, autophagy, calcium homeostasis, and innate immunity [3][4][5][6][7][8][9] . The mitochondrial genome is independent and distinct from that of the nucleus 10 . Human mtDNA, with a few hundred to several hundred thousand copies contained in each cell 11,12 , encodes 37 genes (22 tRNAs, 2 rRNAs, and 13 oxidative phosphorylation protein subunits) in a small circular 16.5 kb double-stranded piece of DNA 13,14 . The exceeded heteroplasmic threshold of mtDNA mutations manifests in multiple disorders ( Fig. 1a) 14 . Specifically, a vast array of more than 250 mitochondrial defects have been implicated in various pathogenic mtDNA variants that primarily affect muscle and nervous tissues 15,16 . Notably, among the 97 known pathogenic mtDNA variants, point mutations accounted for 92 variants (approximately 95%), which indicated transition point mutations, including A > G, T > C, G > A, and C > T, accounting for 87% (MITOMAP: A Human Mitochondrial Genome Database. www.mitomap.org) (Fig. 1b). For instance, the MT-TL1 mutation m.3243 A > G can trigger multiple mitochondrial syndromes, such as chronic progressive external ophthalmoplegia (CPEO); maternally inherited diabetes and deafness (MIDD); and mitochondrial myopathy, encephalopathy, lactic acidosis, and stroke-like episodes (MELAS) syndrome. Meanwhile, three major mtDNA mutations, m.3460 G > A, m.11778 G > A, and m.14484 T > C, are present in more than 95% of patients suffering from Leber hereditary optic neuropathy (LHON), a maternally inherited disease that is associated with a loss of vision 17,18 . Myoclonus epilepsy and ragged-red fiber (MERRF) syndrome, a severe neurodegenerative defect, is predominantly caused by the point mutation m.8344 A > G in the MT-TK gene 19 . Furthermore, a spectrum of mutations in the SPG7 gene, including c.861dupT, c.2221 G > A, c.2224 G > A, c.2228 T > C, c.1672A > T, c.1192 C > T, c.1529 C > T, c.1454_1462del, c.1067 C > T, c.184-3 C > T, c.233 T > A, c.1046insC, c.1457 G > A, and c.1053dup, were revealed in patients suffering from many mitochondrial diseases, such as progressive external ophthalmoplegia (PEO), ptosis, ataxia, spastic paraparesis, cerebellar atrophy, and proximal myopathy 20 . Strikingly, the C5024T mutation in the mtDNA gene MT-TA was uncovered via mouse models of heteroplasmic mitochondrial defects in the smooth muscle surrounding the colon and cardiomyocytes 21 . Although single-base substitutions can theoretically be corrected using base editors, precise and efficient therapeutics for several mtDNA mutations and mitochondriainvolved diseases, have not been reported thus far. The development of and advances in base editing technology for correcting mtDNA base variants promise a potential therapeutic strategy for mitochondrial diseases and the generation of disease models in living cells and animals.
The development of and advances in mitochondrial gene editing tools Research on mitochondrial DNA editing started with the use of restriction endonucleases, such as PstI, SmaI, and ApaLI, engineered to be expressed exclusively in mitochondria (called mitoRes) [22][23][24] . The specific elimination of pathogenic mtDNA is followed by repopulation with normal mitochondrial DNA via heteroplasmy. Since mitoRes cannot be easily redesigned, it is difficult to use them to correct various clinical mutations, as there is the limitation of potentially editing the off-target loci [22][23][24] .
Programmable DNA nucleases, such as clustered regularly interspaced short palindromic repeats (CRISPR)/CRISPR-associated protein 9 (Cas9), which recognize a short sequence known as the protospacer adjacent motif (PAM) and follow the DNA sequence targeted by a functional complex of Cas nucleases and a single guide RNA (sgRNA), has revolutionized gene modification and enabled the editing of mtDNA mutations [25][26][27] (Fig. 2a). Of note, the CRISPR/Cas9 system, the most applied CRISPR construct, has recently been reported to efficiently cleave targeted mtDNA in HEK293T cells and zebrafish 26,27 . However, a key impediment to this approach is the import of the exogenous sgRNA into mitochondria 28 . Although several researchers have shown significant attempts at sgRNA delivery, this as-yet-unresolved limitation demonstrates the inefficiency of this approach 26,29,30 . Recently, one group demonstrated that not only Cas9 (type II) but also Cas12a (one of the type V CRISPR effectors) could access mitochondrial DNA editing 31 . However, although they tried to engineer the guide RNA using the RP loop related to importing mtRNA, clear evidence of efficient Cas12a-mediated mtDNA genome editing was not obtained 31 . Hence, to date, CRISPRbased systems have not been reliably applied to mtDNA manipulation.
Soon after the development of site-specific nucleases such as zinc-finger nuclease (ZFN) and transcription activator-like effector nuclease (TALEN) [32][33][34] , mitochondrially targeted ZFN (mtZFN) and mitochondrially targeted TALEN (mitoTALEN) versions emerged, optimized for delivery capacity into mitochondria, and they have been effectively used for heteroplasmic manipulation in several models of pathogenic mtDNA variants by inducing mtDNA double-strand breaks (DSBs) 35,36 . The sequence specificity of the mtZFN or mtTALEN monomers is demonstrated through interactions between the protein and DNA via tandemly organized repetitive elements derived from either zinc-finger protein (ZFP) or TALE DNA-binding protein (see Table 1 for a summary).
Development of the structure and use of ZFP technology for mtDNA editing Zinc-finger nucleic acid-binding modules are some of the most profuse proteins in eukaryotes and comprise the first programmable nucleic acid-binding domain applied in epigenome engineering [37][38][39] . Each ZFN strategy for nuclear DNA modification includes two DNA-binding domains recognizing and binding a specific DNA sequence and a nonspecific FokI cleavage domain, which is a mediator for DNA excision and fragment insertion or frameshift variants of the DNA target (Fig. 2b). Structurally, the DNA-binding domain, consisting of three ZFP groups (each a 30amino-acid module) bound to a 3-nucleotide DNA sequence, is composed of two adjacent β-sheets and an α-helix that interact with a single zinc atom, whereas the nuclease activity of the FokI cleavage domain, which must contain two distinct yet complementary parts, requires its dimerization 32 . Based on the favorable qualities of the components, the combination of the flexible and specific DNA-binding characteristics of ZFPs and the vigorous ablation ability of the FokI restriction enzyme in the ZFN architecture makes it credible for genome editing. To date, the potential of ZFN technology has been harnessed and effectively   used for mtDNA manipulation. Initial reports revealed a sitespecific methylase status using a three-finger monomeric ZFP equipped with an N-terminal mitochondrial targeting sequence (MTS) and a nuclear export signal (NES) fused to DNA methyltransferase 3a (DNMT3A), specifically targeting the m.8993 T > G mutation 40 . Furthermore, a single ZFP monomer conjugated to two FokI endonuclease domains tethered together via a flexible linker of 35 amino acids was engineered to target m.8993 T > G in heteroplasmic cybrid cells. However, this approach raises a safety concern that the combination of a fourfinger ZFP monomer binding only 12 bp and a constitutively active nuclease might excessively reduce the mtDNA copy number 41 . However, attempts to improve its characteristics are still far from being efficient 42 . Alternatively, optimizing the heterodimeric mtZFN architecture consisting of NARPd(+), an mtZFN specifically binding to a mutant sequence, and COMPa(−), a nonmutant sequence-bound mtZFN monomer, caused an effective shift in the heteroplasmy of cybrid cells bearing the m.8993 T > G mutation and "common deletions" 35 . Changes in the heteroplasmy shifting efficiency have also been demonstrated in a fine-tuned m.8993 T > G model 43,44 . Specifically, approximately 80% of the cybrid cells bearing the m8993T > G mutation were successfully transfected with the specific mtZFN that induced a considerable shift toward the wild-type mtDNA. Strikingly, substantial depletion of the mtDNA copy number was revealed at 24 h, followed by a return to a level comparable to nontransfected cells 28 days post-transfection. Moreover, the model was further capitalized to investigate the metabolic defects in mtDNA mutation-bearing cells 43 . In a later study 44 , the authors exploited a model called mTUNE to produce several isogenic cell lines (mT7, mT45, and mT80) with various levels of m8993T > G mutation (7%, 45%, and 80%, respectively). These m8993T > G heteroplasmic cell types augmented cell migration, a finding supported by reports of MDH1 increasing ATP generation via glycolysis.
Of interest, mtZFNs exert their potential applications in vivo 45 . In particular, a construct containing an m.5024 C > T-specific monomer (MTM25) and a wild-type-specific monomer (WTM1) was produced by encoding the monomers in a distinct adenoassociated virus (AAV) genome, followed by their encapsidation into AAV9.45 vectors subsequently administered to mice 21,45 .
Recently, a ZFP-based base editor for mtDNA called ZFD has been developed for catalyzing the conversion of C to T in human cells 46 . ZFD constructs were created by connecting ZFPs to the split interbacterial toxin deaminase DddA tox and a uracil glycosylase inhibitor (UGI) via linkers (Fig. 2c). The optimized architecture of ZFDs showed successful base editing in HEK293T cells with frequencies of 2.6-30% 46 . Interestingly, the conversion of C to T mediated by ZFDs was performed in most Cs in both the TC and TCC sequence contexts. Although the rates of C to T conversion are currently low, further studies on engineering ZFDs to enhance both efficiency and precision may open a new door for correcting pathogenic mtDNA mutations in human embryos, fetuses, and specific adult tissues 46 .
Development of the structure and use of TALEs technology for mtDNA manipulation MitoTALENs have been platform-engineered via the fusion of the DNA-binding domain of TALEs, comprising monomers from the plant pathogenic bacteria genus Xanthomonas, and the nonspecific cleavage domain of FokI endonucleases 47,48 . The DNAbinding domain that recognizes the specific nucleotide in the target DNA sequence consists of tandem conserved repeats of 33-35 amino acids (Fig. 2d). Among them, two positions (residues 12 and 13), known as the repeat variable diresidue (RVD), show high variation in the recognition ability for a specific nucleotide 49 . Notably, 4 out of 25 RVDs (HD, NI, NG, and NN), which were the most prevalent, could identify C, A, T, and G, respectively 50 . In contrast to ZFNs, the structural foundation of the TALE DNAbinding specificity is more amenable to engineering and requires a minimal number of RVD panels encoding all base specificities 51 .
The first study on the development of mitoTALENs demonstrated the elimination of common deletions and the m.14459 G > A point mutation in mtDNA by applying two engineered constructs: Δ5-mitoTALEN and 14459A-mitoTALEN, respectively 36 ; these architectures cause heteroplasmic shifting in heteroplasmic cybrid cells. Interestingly, a design based on the requirement of a T at position 0 in the sequence responsible for DNA recognition, called the T 0 strategy, was used for m.13513 G > A mutation elimination in heteroplasmic cybrid cells 52 . However, attaining the binding specificity for a single nucleotide at a mutation position that is neither N > A nor N > T is currently far from easy. Furthermore, the T 0 strategy requires two monomers, thus limiting the packaging capacity in various vectors, especially AAVs. To improve these hurdles, the authors showed that a shortened mitoTALEN monomer could also be applied for the efficient elimination of m.8344 A > G mutations. In addition, later studies depicted changes in the heteroplasmic shifting efficiency in iPSCs harboring m.13513 G > A and m.3243 A > G 53,54 . Treatment with mitoTALENs targeting the CD 5' region of mtDNA was shown to induce DSBs, resulting in a common deletion, followed by an accumulation of mtDNA 49,55,56 . In addition, human LHON m.14459 G > A and NARP m.9176 T > C mutations were specifically eliminated by injecting mitoTALENs into MII stage mouse oocytes 57 .
Interestingly, a monomeric variant of mitoTALENs known as mitoTev-TALEs, which are engineered by fusing a TALE-binding domain with a T4 phage-derived I-TevI homing endonuclease catalytic domain through a linker, has been used to manipulate m.8344 A > G mutant mtDNA in heteroplasmic cybrids 58 . However, this architecture has apparent disadvantages, such as requiring a CNNNG cleavage site for I-TevI.
To further explore the potential of TALE, hybrid platforms have been constructed by the fusion of solely mtDNA-binding TALEs and base-editing deaminases for precise modification of mtDNA. Recently, DdCBEs that can precisely and specifically convert C to T in human mtDNA have been formed by adding TALE proteins and UGI to split-DddA (Fig. 2e) 59 . Combining two TALE proteins with the mtDNA target resulted in the activation of the inactively nontoxic split-DddA halves. Surprisingly, the addition of UGI, protecting U from base excision by glycosylase, improved the efficiency of base editing by approximately 8-fold. Most excitingly, DdCBEs have been successfully exploited for mtDNA base editing in various species, such as human embryos, zebrafish, mice (embryos and adults), rats, and plants [60][61][62][63][64][65][66][67][68] .
Although the discovery of DdCBEs is a huge step forward, this technology is limited by the dependence of DddA on editable sites where Cs in the 5'-TC-3' contexts are preferably converted to Ts (conversion from 5′-TCC-3′ to 5′-TCT-3′ or 5′-TTT-3′ is also available depending on the target). As a consequence, methods based on the original DdCBEs could correct only 10% of pathogenic mtDNA point mutations 69 . Soon after, the authors developed DddA variants using rapid phage-assisted continuous evolution. DddA6-containing DdCBEs conduct C•G to T•A conversion at TC sites 3.3-fold more efficiently than canonical DdCBEs. Furthermore, base editing with the DddA11 variant revealed strong compatibility over a range of settings (TC, AC, and CC), in which the editable levels for CC and AC in both the nucleus and mtDNA were increased by 15-30% 70 . Despite effective baseediting improvements, engineered DdCBEs are still limited to only being able to edit C-to-T residues. However, this limitation has been circumvented by the advent of TALEDs, a novel tool that enables the conversion of A-to-G 69 . This platform consists of an mtDNA-targeting TALE, adenosine deaminase TadA8e, and inactive DddA tox as a cytosine deaminase (Fig. 2f). Moreover, TALEDs have been shown to successfully correct 47% of known pathogenic mtDNA point mutations in human cells 69 . More intriguingly, TadA8e, which supposedly operates on singlestranded DNA, was shown to possess an unexpected capability of adenine deamination in double-stranded DNA. This observation was explained by the presence of DddA tox , which enables the unwinding of the double-stranded DNA and provides TadA8e superfast access to the DNA for the necessary modifications 69 . However, this technology has the limitation of bystander editing, which converts nucleotides adjacent to the desired targets in the same editing window. Optimization in the engineering of DddA tox or TadA might eliminate this undesired editing 69 . In addition, it would be intriguing to enhance the specificity and efficiency of TALEDs, which might provide further possibilities for mtDNA editing. Novel methods for the TALED-mediated manipulation of ES cells or direct editing in mouse embryos are certain to be developed in the near future.

Off-target effects by mitochondrial gene editing tools
Solving unintended editing problems is of foremost importance for safe gene editing applications in clinical therapies. In the utilization of programmable editing platforms such as mtZFN and mitoTALEN, attenuating the temporary total mtDNA copy number is of paramount concern due to the depletion of mutant mtDNA before the repopulation of the wild-type mtDNA. Notably, the depletion might be significantly affected by the off-target editing of mtDNA as a consequence of inadequate reagent specificity or inappropriate concentrations of cleavage agents. Although offtarget events of mtDNA editing using DdCBEs are relatively rare in HEK293T cells 59 , this platform can cause considerable mtDNA offtarget editing in zebrafish and plants 62,65,66,68 . Most strikingly, two recent studies using DdCBEs have elucidated substantial off-target editing of the nuclear genome and mtDNA in mammalian cells and mice 71,72 . For instance, DdCBEs preferably caused undesired low-frequency mtDNA editing events at 5′-TC-3′ sequences, which were most prominent in the spacer regions to the left and right of the TALE binding sites, with some lying outside the regions 71 . Most unexpectedly, when delivered to fertilized 2-cell stage embryos by the genome-wide off-target analysis by two-cell embryo injection (GOTI) method, DdCBEs also induced remarkable sequence-independent off-target effects, resulting in singlenucleotide variants (1000-1500) in the nuclear DNA. The unexpected results are likely due to the unique characteristics of the DddA tox cytosine deaminase used in DdCBE, which favors dsDNAs as substrates, unlike the substrate preference of cytosine deaminase APOBEC1 in the BE3 protein for ssDNA 71 . Further studies will be necessary to clarify the effects of various MTSs on off-target editing events by DdCBEs as well as to develop novel approaches to reduce the adverse effects. Using an unbiased method (Detect-seq) to evaluate genome-wide specificity, Lei. et al. 72 demonstrated that hundreds of nuclear off-target edited sites were induced by DdCBEs in both TALE array sequence-dependent and sequence-independent cases. The studies suggested that certain interactions between the DdCBE and the CTCF binding region seem to be associated with sequence-independent off-target effects. However, the exact mechanism underlying this observation remains to be determined. Interestingly, this offtarget editing issue was significantly improved by applying several advanced DdCBE constructs: UGI-NES-DdCBE, to which NES sequences were added to hinder the localization of DdCBE in the nuclei, and DddI A -DdCBE, in which DddI A , a natural immunity protein of DddA, was used to preclude nuclear DNA editing by DdCBE by obturating the active center of DddA tox 59,72 . Nonetheless, further thorough analyses of off-target editing of both mtDNA and nuclear DNA by various mtDNA editing technologies will be necessary for both basic research and clinical applications in the future.

CONCLUSION
The potential of ZFPs and TALEs is promising, and their approaches for the base editing of mtDNA have been demonstrated. However, these platforms are limited by size compatibility, which makes their delivery to mitochondria less efficient using viral systems such as AAVs. In addition, there is a need for more efficient design and assembly methods for mtZFN and mitoTALEN monomers to recognize a wide spectrum of mtDNA sequences 52,73 . Consequently, the production of these proteins requires cost-and labor-intensive processes. These challenges could be overcome by using CRISPR/Cas9, a powerful editing system applicable to a wide number of organisms, including plants and mammals. CRISPR/Cas9 relies on only two components: (i) sgRNA, which recognizes a specific DNA sequence, and (ii) Cas9 nuclease, which cleaves the DNA sequence. However, as mentioned above, a major challenge in applying this system in mtDNA editing is to find effective methods for delivering exogenous sgRNA into the mitochondrion. If this could be achieved, the CRISPR/Cas9 system could potentially be broadly applied to manipulate mtDNA.
In fact, several stem-loop motifs mediating the shuttling of nuclear RNA molecules into mitochondria have been characterized at the 5′ or 3′ ends of a variety of endogenous RNAs, such as 5S rRNA, RNase MRP, and H1 RNA, which play an essential role in the functioning of mammalian mitochondria [74][75][76] . In particular, stem-loop motifs from H1 RNA, which are appended to cytosolic RNA, facilitate the efficient importing of the RNA molecules into mitochondria 56 . Furthermore, it has been suggested that some 5S RNA motifs and two domains (yeast cytosolic tRNALys (CUU)derived F-arm and D-hairpin) can also assist the uptake of synthetic RNA into mitochondria [77][78][79] . Moreover, it has been reported that the fusion of an MTS with the AAV2 capsid protein VP2 helped to target the NADH-encoding ND4 gene in mitochondria 80 . Strikingly, a recent study revealed that the Fig. 3 Strategies for delivering gene editing platforms into mitochondria. ZFP-and TALE-based gene editors or engineered CRISPR/Cas systems modified with MTS and stem-loop motifs can be imported into mitochondria via transfection with AAVs or eVLPs, which encapsulate plasmid DNA or mRNA encoding the gene editing platforms. mtZFN, mitochondrial zinc-finger nuclease; TALEN transcription activator-like effector nuclease, DdCBE DddA-derived cytosine base editor, ZFD zinc-finger deaminase, TALE-linked deaminase (TALED); CRISPR/Cas clustered regularly interspaced short palindromic repeats/Cas. use of an engineered Cas9 nuclease linked with N-or C-terminus MTSs for editing mtDNA mutations in HEK293T cells and zebrafish resulted in a significant reduction in mtDNA copy number 27 . These reports provide a critical proof of concept for the delivery of sgRNA appended to a stem-loop element into mitochondria, followed by an interaction with the Cas9 nuclease and cleavage of specific mtDNA sequences. Moreover, not only Cas9 but also Cas12a can enter the mitochondria by using MTSs, and there have been attempts at mitochondrial delivery by engineering various stem-loop motifs on guide RNAs 31 . Together, the studies suggest that efficient delivery of sgRNA via stem-loop motifs show great promise for mtDNA-editing systems (Fig. 3).
Recently, a transient ribonucleoprotein (RNP) delivery platform using engineered virus-like particles (eVLPs) has been reported 81 (Fig. 3). Essentially, eVLPs are empty viral shells into which therapeutic RNPs can be packed and delivered to the target DNA with minimal off-target editing effects 81 . Specifically, a newly engineered variant of eVLPs (v4 eVLPs) showed 16-and 4.7-fold increases in the packaging capacities of base editor RNPs and Cas9, respectively, compared to earlier architectures 82 . Most excitingly, eVLPs have been successfully used for delivering therapeutic packages to various organs in mice (liver, brain, and retina). These studies underscore the possibility of utilizing eVLP to deliver gene-editing cargoes for mtDNA to cells or multiple targeted organs in both animals and humans with minimal off-target editing 81 . Despite the potential of transient and efficient delivery of eVLPs, further studies on their pharmacokinetics, such as half-lives, are imperative. Additionally, determining the adequate requirements for base-editing cargoes may be a genetic therapy milestone for mtDNA editing in the future. Recently, extraordinarily novel applications based on both TALEs and ZFPs have been developed for highly specific mtDNA editing platforms toward heteroplasmic shifting with minimal cytotoxicity.
Optimizing the engineering of these platforms will surely pave the way for genetic approaches toward clinical trials by reducing the size to improve the packaging capability of viral systems and promoting efficiency and specificity for mtDNA manipulation. In fact, novel types of CRISPR systems, mini CRISPR-AsCas12f1 and miniature CRISPR-SpaCas12f1, consisting of 422 and 497 amino acids, respectively, can be feasibly and readily delivered to bacteria, cells, and tissues employing various delivery methods, such as AAV, plasmid, and RNP, have been reported 83,84 . Due to the advantages of editing efficiency and delivery, miniature platforms should also be considered an effective approach for mtDNA editing. Although the successful delivery of exogenous sgRNA into mitochondria using CRISPR-Cas systems could prompt a revolution in mtDNA editing, it is undeniable that there are still real-world obstacles that need to be resolved for exceptional competence in the application of TALEs and ZFPs.