In vivo gene editing of outer hair cells prevents progressive hearing loss in a dominant-negative KCNQ4 murine model

Adult-onset hearing loss (AHL)—including presbycusis—caused by outer hair cell (OHC) degeneration is the most common sensorial disorder. Despite the high prevalence of AHL and wide therapeutic window, no targeted therapy is currently available. Here, we generated a mouse model harboring Kcnq4 W276S/+ to recapitulate DFNA2, a common genetic form of progressive hearing loss caused by degenerating OHCs. By comprehensively optimizing guide RNAs, Cas9s, vehicles, and delivery routes, we found that in vivo gene editing using dual adeno-associated virus packaging in OHCs via the round window membrane significantly improved auditory function. We developed a new technique using live-cell imaging to measure the membrane potential of the OHCs and demonstrated that our approach resulted in more hyperpolarized, steady-state OHCs, indicative of elevated KCNQ4 channel activity. These findings can help develop targeted therapy for AHL and support the use of CRISPR-based gene therapy to rectify defects in OHCs.


Introduction
Hearing loss is a common sensory disorder 1 AHL can develop due to many risk factors (genetic and environmental) 5 . Lifetime acoustic noise exposure is often the primary cause of age-related hearing loss 6 . Moreover, alterations in genes that encode potassium channels located in the cochlea-the main component associated with the development of AHL-can influence individual susceptibility to noise exposure 7,8 . The leading cause of AHL is the degeneration of outer hair cells (OHCs), often triggered by genetic vulnerability to noise exposure 6,9 . In this respect, functionally restoring OHCs is indispensable for preventing and treating AHL.
Genome editing technologies, including Cas9 nucleases, base editors, and prime editors, are emerging as tools for treating genetic deficits responsible for causing various diseases 10,11 . In particular, Cas9 nucleases are effective and useful for disrupting dominant-negative alleles, whereas base and prime editors are potentially useful for rescuing recessive loss-of-function mutations. While gene replacement or silencing therapies for hearing loss have been attempted in several mouse models [12][13][14][15][16][17][18][19] , in vivo gene editing faces numerous therapeutic hurdles, including the selection of suitable vehicles and delivery routes to efficiently and safely deliver the gene editing system 20,21 . Previously, in vivo gene editing with Cas9 nuclease or base editor had been employed in mice with transmembrane channel-like 1 gene (Tmc1) variants to restore hearing [22][23][24] . However, previous in vivo gene editing techniques have not provided sufficient transduction rates in OHCs; therefore, in vivo gene editing systems need to be further optimized before their application in a clinical setting 25 .
This study attempted to employ in vivo gene editing to treat AHL resulting from OHC degeneration using a newly established murine model with a dominant-negative effect of a missense mutation (p.W276S) in the voltage-gated potassium channel subfamily Q member 4 gene (Kcnq4). Moreover, we comprehensively optimized the efficiency, safety, and delivery route of the Cas9 nuclease system to improve the feasibility of using in vivo genome editing to treat AHL.

Selection of KCNQ4 for in vivo gene editing
We first attempted to identify genetic predispositions that contribute to the development of AHL.
To this end, we enrolled 213 patients with AHL and performed whole-exome or targeted deafness panel sequencing. The results showed that genetic mutations accounted for 33.8% of AHL, and KCNQ4 was the most frequently mutated gene in this cohort ( Fig. 1a and Supplementary Table 1). Indeed, DFNA2 (deafness, autosomal dominant type 2) associated with KCNQ4 mutations is one of the most common causes of autosomal dominant hearing loss 26,27 . We selected KCNQ4 as a target for in vivo gene editing based the following reasons: 1) KCNQ4 is also associated with age-related hearing loss and noise-induced hearing loss 8,28 ; 2) DFNA2 usually exhibits progressive hearing loss, which provides a broad therapeutic window; and 3) though KCNQ4 is exclusively expressed at the basolateral surface of OHCs (Fig. 1b), it is indispensible for maintaining the electrolyte milieu of the endolymph via potassium recycling in the cochlea 29 .
Subsequently, knock-in mice harboring the Kcnq4 c.830G>C allele and a silent mutation at nucleotide position 810 were generated (Fig. 1c). The silent mutation (c.810C>A; p.S269=) not only facilitated genotyping by making only the mutant allele susceptible to Nde1 but also discriminated between the edited mutant alleles containing the indel and the wild-type alleles affected by Cas9-single guide-RNA (sgRNA) in in vivo gene editing experiments.
We found that Kcnq4 W276S/W276S (homozygote) and Kcnq4 W276S/+ (heterozygote) mice distintly exhibited moderate hearing loss in the auditory brainstem response (ABR) at P21 (postnatal 21 d) (Fig. 1d). A progressive hearing loss phenotype was also validated in the followup hearing function tests performed at P49 and P77. Hearing loss was attributed to the degeneration of OHCs, particularly in the high-frequency region in the cochlea ( Fig. 1e and f).
The morphology of stereocilia in the surviving hair cells of Kcnq4 W276S/+ mice was normal (similar to that in wild-type mice; Supplementary Fig. 1). Hearing thresholds at all frequencies reached almost scale-out levels at P49, which was determined as the time point of deep sequencing in the subsequent experiments (Fig. 1d)

Optimizing the targeting efficiency of the Cas9-sgRNA system for Kcnq4 c.830G>C
To select active and efficient sgRNA for target Kcnq4 mutation (c.830G>C) editing, all possible sgRNA target sequences with a protospacer-adjacent motif (PAM; 5′-NGG-3′) downstream of the mutation locus were designed ( Fig. 2a and Supplementary Fig. 2a). The gene editing scores for each sgRNA were determined using the deepSpCas9 prediction program 33 , which confirmed that the combination of SpCas9 (Strepotococcus pyogenes) wild type and T3 was the most efficient among the various types of SpCas9 (Supplementary Table 2). To elucidate the activity of these sgRNAs, we generated a reporter cell line expressing the Kcnq4 mutation (c.830G>C) target sequence with a dual-fluorescent reporter construct. The presence of the reporter was detected by a sole RFP signal, whereas nuclease activity was analyzed by simultaneous GFP and RFP fluorescence ( Fig. 2b and Supplementary Fig. 2b). Next, we isolated the genomic DNA from reporter cells in the transfected pooled cells and performed T7E1 assay and NGS analysis to evaluate the indel frequency at target sites. The mutation frequencies at the target sites ranged from 31-34% and 20-24% as per the T7E1 assay and NGS, respectively (Fig. 2c).
To determine the wild type off-target cleavage activity of sgRNA, genomic DNA from wild-type Neuro2A cells transfected with SpCas9 and individual sgRNA expressing plasmids were analyzed using the T7E1 assay. We observed indel frequencies of 9%, 41%, 0%, and 47% for each respective sgRNA (T1-T4) in the wild type target sites (Fig. 2d). Finally, sgRNA-T3 was selected for subsequent in vivo gene editing because it showed no cleavage activity for the wild-type target locus while exhibiting high-cleavage activity for the c.830G>C mutation target sequence (Fig. 2c, d).
Next, we optimized both adeno-associated virus (AAV) and ribonucleotide complex (RNP) materials for in vivo intracochlear delivery. For AAV packaging, we chose the AAV2/Anc80L65 serotype due to its high efficiency in the OHCs 34,35 . We designed split Cas9 and gRNA-T3 into two plasmids (Supplementary Fig. 3a-d). The final injected AAV titer was estimated to be 1.00 × 10 9 genomic copies in one cochlea with a volume of 1 µL ( Supplementary Fig. 3e). For RNP packaging, we optimized the liposome packaging protocol using Lipofectamine 2000 as it has been reported to have high efficacy for RNP application in the cochlea (Supplementary Fig. 4a) 22 . After analyzing various titer combinations of the SpCas9 protein, sgRNA, and Lipofectamine 2000, we determined the optimal RNP mixture ratio and obtained up to 32.2% indel efficiency under in vitro conditions (Supplementary Fig. 4b).
To determine the best delivery route, we investigated the in vivo efficiency of viral injection through the round window (RW), scala media (SM), posterior semi-circular canal (PSCC), and utricle using AAV2/Anc80L65-eGFP (Supplementary Fig. 5a) 36 . The SM route via cochleostomy or RW route showed the highest eGFP expression levels in the OHCs across all frequency regions (Supplementary Fig. 5b). Therefore, we mainly injected RNP into the SM via cochleostomy or AAV into the RW. Although cochleostomy damaged the hearing function in some cases due to technical difficulties, in most cases hearing loss was not induced in wild-type mice when cochleostomy was appropriately performed (Supplementary Fig. 6).

Phenotypic correction of Kcnq4 p.W276S mice by CRISPR/Cas9 gene editing
After injecting the appropriate AAV and RNP packages into Kcnq4W276S/+ mice within P4 d, phenotypic rescue was evaluated by ABR and distortion-product otoacoustic emission (DPOAE) at 3 and 7 weeks after injection (Fig. 3a). The representative traces of ABR waves 7 weeks after injection are shown in response to varying sound intensity of 6-kHz tone bursts (Fig. 3b). In the AAV-injected ears, the ABR threshold, P1 amplitude, and P1 latency were significantly improved at all frequencies compared with those of uninjected ears ( Fig. 3c and Supplementary   Fig. 7a-c). The amplitude in DPOAE in the AAV-injected ears was significantly increased at frequencies corresponding to the 12-and 16-kHz regions (Fig. 3d). Meanwhile, in the RNPinjected mice, improvements in ABR threshold, P1 amplitude, and P1 latency were comparable to or less than those in the AAV-injected mice ( Fig. 3e and Supplementary Fig. 7d-f). However, the amplitudes obtained by DPOAE in the RNP-vehicle-injected ears were not significantly improved, except in the 12-kHz region (Fig. 3f). Taken together, both AAV-and RNP-vehicle injection effectively restored hearing function in Kcnq4 W276S/+ mice.

Gene editing efficiency in phenotypically corrected mice
Gene editing efficiency was evaluated using allele-specific indel analysis of phenotypically corrected mice. When the best edited in vivo sample was compared with an ex vivo sample, only 0.6% mutant alleles were corrected in the AAV-injected case, showing diverse indel patterns (similar to ex vivo samples, 1.5%; Fig. 4a). Nevertheless, reproducible and relevant sequenced fragments having out-of-frame insertion and valid deletion patterns at the target cleavage site confirmed the validity of genotype correction. Furthermore, the fold changes in the editing rates correlated with the threshold changes in ABR after AAV and RNP injection (Fig. 4b).
When gene editing efficiency rates were matched with the degree of hearing improvement (categorized as marked, moderate, mild, and minimal groups) based on the average differences in ABR thresholds between the injected and contralateral uninjected cochlea, significant correlation was observed in marked, moderate, and mild groups in the AAV-injected cases (Fig. 4c). In the RNPinjected cases, only the marked and moderate groups exhibited significant differences. Overall, the gene editing efficacy of the AAV-vehicle for rescuing hearing loss was more favorable and higher than that of the RNP-vehicle. Therefore, we investigated the mechanism responsible for this distinct phenotypic rescue in the AAV-injected mice.

Effect of CRISPR/Cas9-induced ablation of mutant alleles on cochlear function
We found that the viability of OHCs in the Kcnq4 W276S/+ mice with improved hearing thresholds after AAV injection was not significantly different from that of OHCs in uninjected mice (Fig. 4d, e and Supplementary Fig. 8a). Similarly, there were no differences in KCNQ4 expression in the OHCs (Supplementary Fig. 8b), the extent of neurofilament innervated into hair cells ( Supplementary Fig. 8c, d), and the number of spiral ganglion neurons (Supplementary Fig.  9a, b) between the AVV-injected and uninjected groups 7 weeks after injection. Finally, no morphological difference was observed in the stereocilia of the IHCs and OHCs between the two groups ( Supplementary Fig. 9c). Therefore, hearing restoration by the CRISPR/Cas9 system could not be explained by the survival of hair cells and neurons. Thus, we hypothesized that the surviving cells functionally differ with respect to mutant allele ablation. Therefore, we developed a novel ex vivo hair cell thallium imaging technique to reflect the state of membrane potential. Because the potassium (or thallium) influx into the hair cells via nonselective cationic channels is affected by the electrochemical gradient, the ion influx amount depends on the membrane potential. Therefore, more hyperpolarized cells are associated with faster thallium influx (Supplementary Fig. 10). Given that KCNQ4 in the basolateral surface can hyperpolarize hair cells 29 , we speculated that the restored activity of KCNQ4 by gene editing increases the influx of thallium ions through apical cationic channels. To evaluate the electrophysiological rescue of KCNQ4 by AAV injection, the uncapped cochlea incubated with FluxOR dye was visualized using fluorescence microscopy (Fig. 5a). We successfully observed that the thallium influx was exclusively dominant in the OHCs of Kcnq4 +/+ mice ( Fig. 5b and Supplementary Movie 1). In addition, the influx of the thallium ions depended on the leaky activity of mechano-electrical transduction (MET) channels at the apex during the resting steadystate (Fig. 5c). Because we did not induce mechanical stimulation to activate the hair cells, thallium influx solely depended on the membrane potential of the hair cells, regardless of the activation state of MET (i.e., TMC1) or KCNQ4 channels [37][38][39] .
FluxOR signals increased more rapidly in the OHCs of mice with AAV injection than in those with mock injection (Fig. 5d). Both the OHCs and IHCs showed a significantly increased thallium influx slope after AAV injection (Fig. 5e, f). These findings indicate that the surviving hair cells in the Kcnq4 mutant mice are more hyperpolarized, similar to those in the wild-type mice after the mutant allele ablation by gene editing.

In vivo gene editing of KCNQ4 variants linked to DFNA2
To investigate the clinical feasibility of in vivo gene editing to reverse the hearing loss caused by variants in KCNQ4, we searched all the variants in KCNQ4 associated with DFNA2 from the databases, including the Human Gene Mutation Database (HGMD professional v2020.4) and ClinVar. We found 49 mutations in KCNQ4, including frameshift and splicing variants, and then analyzed the best gene editing scores for each mutation using the deepSpCas9 prediction program with various combinations of sgRNA and Cas9 variants 33 . We found that the gene editing scores for most mutations were comparable to or even higher than those for p.W276S ( Fig. 6 and Supplementary Table 3). In particular, wild-type SpCas9 was the most efficient for all mutations in KCNQ4. This indicates that the majority of variants in KCNQ4 may be potential targets for in vivo gene editing.

Discussion
In the present study, pathogenic KCNQ4 variants were the most commonly identified in patients with AHL. We successfully applied in vivo gene editing with Cas9 nuclease to ameliorate progressive hearing loss in the dominant-negative Kcnq4 W276S/+ murine model. Moreover, we revealed that even minimal gene editing efficiency can restore the hyperpolarized steady-state of OHCs in the cochlea.
While in vivo genome editing can remedy an extensive range of genetic diseases, the cochlea-targeting capabilities need to be further improved. Compared with delivery to organs such as the liver, muscle, brain, and eye, the delivery of gene-editing-related materials to the cochlea is very difficult 40 as the cochlea is surrounded by a hard cortical bone and bloodlabyrinth barrier that hinders efficient delivery. However, our results indicate that strong gene editing efficiency is not necessarily required to sufficiently restore hearing in the Kcnq4 W276S/+ murine model; specifically, ~0.6% of gene editing efficiency at the genomic DNA level sufficiently rescued the auditory phenotype with dual AAV plasmids with split SpCas9 and sgRNA. These data are consistent with those of a previous study, which reported sufficient hearing restoration when the in vivo gene editing efficiency was ~0.6% in Beethoven mice harboring the Tmc1 mutant allele 41 . The discrepancy between the efficiency and phenotypic rescue can be explained by several technical reasons. For instance, following the preparation of cochlear samples for sequencing, the actual number of inactivated mutant alleles may be inevitably underestimated in some cases. In addition, genetic mosaicism arising from CRISPR-Cas9-mediated non-homologous end joining (NHEJ) and homology-directed repair is often observed in vitro and in vivo 42 ; therefore, the effect of Cas9 on the mutation cut site may be compromised. Nevertheless, we consider that the gene editing threshold required for observable hearing restoration in the cochlea is relatively lower than that for other diseases.
KCNQ4 is crucial for the pathogenesis of AHL, as demonstrated by previous studies 6 .
Genetic alterations in KCNQ4 result in the absence of potassium recycling in the OHCs of the cochlea, a phenomenon that enhances susceptibility to noise and promotes progressive hearing loss. In this respect, genome editing to treat AHL should focus on the OHCs, which are the sites where this sensory organ begins to degenerate. Interestingly, the degeneration of OHCs occurs slowly for decades with onset occurring in older individuals, suggesting that the therapeutic window for treating AHL using gene therapy is adequate for attempting to decelerate progressive hearing loss 26,43,44 . Given that most mutations related to DFNA2, including indel mutations, can be efficiently edited via SpCas9 and its variant forms (Fig. 6), we theorize that this gene editing strategy, i.e., inactivating KCNQ4 dominant-negative mutant alleles, can be applied to other KCNQ4 variants as well.
Although these data indicate that in vivo gene editing is applicable for treating AHL caused by the degeneration of OHCs, the efficiency of gene editing and the delivery vehicle used should be further improved. The high-fidelity capsid of AAV, which is optimized for targeting the OHCs, will be required to accomplish these goals; this strategy has improved the transduction rate of AAV9-PHP.B capsid into OHCs 41 , and can therefore be used for further gene editing applications. Moreover, other delivery methods, such as via mRNA and RNP, will be considered to avoid the collateral safety issues associated with administering viral injections to humans. Although RNP injection aids in correcting alleles in the cochlea, as seen in the present and previous studies 22 , the administration of RNP should be more comfortable than performing a cochleostomy, which is still a challenging procedure for many otology surgeons.
In conclusion, we demonstrated that editing of the Kcnq4 gene in the OHCs-in an effort to enhance the functional channel activity of KCNQ4-can sufficiently restore hearing function in a murine model that recapitulated DFNA2, even at low gene editing efficiencies. Our findings provide a rationale for the clinical application of gene editing to treat AHL.  43 . genotyping was performed using PCR with the same primers, and the PCR products were digested using Nde1 for 2-3 h to identify the two-band pattern in agarose gel electrophoresis.
The plasmids were purchased from Toolgen (Seoul, Republic of Korea); the vector maps of these plasmids are shown in Supplementary Fig. 3.
For the AAV experiments, the previously validated expression plasmids were used to express the split C-Intein-C-Cas9 expression plasmid (N-Cas9 N-intein and C-Intein C-Cas9 plasmids, a gift from Oskar Ortiz) 45 . To prepare the sgRNA-N-Cas9-N-intein, we modified the backbone plasmid (addgene plasmid # 60958) generated by Swiech et al. 46 . Briefly, the backbone vector was digested with XbaI and EcoRI, and the N-Cas9-N-intein fragment digested by XbaI and EcoRI was ligated to the backbone (addgene plasmid # 60958) vector to express the U6-sgRNA-N-Cas9 N-intein. Finally, to clone the selected sgRNA (sgRNA-T3), the vector was digested with SapI and ligated with annealed oligonucleotides. All plasmid and oligonucleotide sequences are shown in Supplementary Table 4 and Supplementary Note 1.

sgRNA preparation and transfection of reporter cell line
The sgRNA target sequences were manually designed based on the protospacer-adjacent motif (PAM-NGG) sequences near the mutation target locus (c.830G>C) ( Fig. 2a and Supplementary   Fig. 2a). The sgRNA-encoding vector (pRG2-sgRNA) was digested with BsaI and ligated with the annealed oligonucleotides; oligonucleotide sequences are listed in Supplementary Table 4. The reporter vector was designed based on a previously reported comparable construct 47,48 . The Lenti-reporter plasmid constitutively expresses RFP and the respective Kcnq4 target sequence containing a c.830G>C mutation (35 bp in length), along with additional sequences encoding eGFP, positioned out-of-frame relative to RFP (Fig. 2a and Supplementary   Fig. 2b). The target sequence contains a 20-bp region with adjacent PAMs for gRNA binding and Cas9-mediated double breaks (Supplementary Fig. 2b). Thus, the recruitment of Cas9 to the respective Kcnq4 mutant-specific target sequence by properly spaced sgRNA results in a double-stranded break, leading to an insertion or deletion (indel) mutation at the target locus.

Reporter cells for the
Because of the codon triplet, one out of three repairs results in a significant "in frame" fusion of eGFP with mRFP located upstream of the cleavage site.

IVT-sgRNA preparation
sgRNA was transcribed in vitro using T7 RNA polymerase with templates generated by annealing and the extension of two complementary oligonucleotides (Supplementary Table 4) with a mMESSAGE mMACHINE® T7 Ultra RNA Synthesis kit (#AM1345; Invitrogen, Carlsbad, CA) according to manufacturer's instructions. RNA was purified using a MEGAclear kit (#AM1908; Invitrogen). Purified RNA was quantified using a Nanodrop and gel electrophoresis.

Lentivirus production and reporter cell line generation
Lenti-reporter plasmids containing the Kcnq4 mutation (c.830G>C) were constructed as previously described 47,48 . Briefly, oligonucleotides, including the target sequence (Supplementary Table 4), were synthesized by Macrogen and annealed in vitro using a thermocycler (95 °C for 5 min and then ramped down to 25 °C at 5 °C/min). The annealed oligonucleotides were ligated into the Lenti-reporter vectors digested with EcoR1 and BamH1.
The lentivirus was produced as previously described 33,49,50 . Briefly, three transfer plasmids, containing the Kcnq4 mutation, psPAX2, and pMD2.G, were mixed at a weight ratio Next, a reporter cell line was generated as previously described 33,49,50  U/mL penicillin, 100 µg/mL streptomycin, and 10% fetal bovine serum.

T7E1 assay
T7E1 assay was performed as previously described 52,53 . Briefly, the target site was amplified using nested PCR with the appropriate primers (Supplementary Table 4). Amplicons were denatured by heating and annealed to allow the formation of heteroduplex DNA and treated with T7 endonuclease 1 (5 U; New England Biolabs) for 20 min at 37 °C, followed by electrophoresis on a 2% agarose gel. Mutation frequencies were calculated as previously described based on the band intensities using ImageJ and the following equation 54 : mutation frequency (%) = 100 × (1 − (1 − fraction cleaved) 1/2 ), where the fraction cleaved is the total relative density of cleavage bands divided by the sum of the relative density of the cleavage bands and uncut bands.

DNA extraction and targeted deep sequencing
The extracted genomic DNA of HEK293T cells was isolated using the Wizard Genomic DNA

Allele-specific indel analysis
To analyze the genotype correction rate by CRISPR on mutant alleles, allele-specific indel analysis strategies were utilized. Gene editing efficiency of mutant alleles, rather than that of wild-type alleles, was prioritized. Because indels at the mutation site can hinder identification of the origin of the edited allele (i.e., whether the wild-type or mutant allele was edited), the mutant allele was identified using the synonymous variant c.810C>A (Fig. 1b)

AAV vector generation
Plasmids of split Cas9 (i.e., C-Cas9 and N-Cas9) and gRNA were packaged into AAV/Anc80 using the Harvard vector core 34 . AAV titers were validated using qRT-PCR targeting the inverted terminal repeat of the virus (Supplementary Fig. 3e). AAV stock concentrations were 1.29 × 10 12 and 4.18 × 10 12 genome copies/mL for C-Cas9 and N-Cas9 with sgRNA, respectively. The final injected AAV titer was estimated to be 1.00 × 10 9 genome copies/1 µL in one cochlea.

Optimization of RNP injection material using in vitro and explant samples
To identify the optimal injection-mixture ratio and incubation times, three injection routes were evaluated as described previously with slight modifications (Supplementary Fig. 4a). Purified  Supplementary Fig. 4b).
As the practically available volume of injection materials into the cochlea of the P2-P5 pup is within 1 µL, the sgRNA (1.0 µg) and Cas9 (1.5 µg) concentrations were maximized.

Inner ear injection
AAV virus or the RNP complex were injected into the inner ear of Kcnq4 W276S heterozygous mutant pups at P1-P3. After the pups were anesthetized by exposure to ice for 2 min (hypothermia), a post-auricular incision was made to expose the injection route. Using a glass pipette and a Nanoliter2020 Injector (World Precision Instruments, Hertfordshire, UK), the injection material (1 µL) was delivered into cochlea at a consistent rate of 40 nL/min. After injection, a suture was made to close the cut skin. The pup was placed on a heating pad to recover for at least 5 min. Various injection routes were comprehensively tested in more than 500 pups to minimize the physical damage to the cochlea during the injection.

Hearing test including ABR and DPOAE
ABR thresholds were measured in a sound-proof chamber using Tucker-Davis Technologies (TDT) RZ6 digital signal processing hardware and BioSigRZ software (Alachua, FL, USA).
Sub-dermal needles (electrodes) were positioned at the vertex and ventrolateral to the right and left ears of anesthetized mice. A calibrated click stimulus (10-µs duration) or tone burst stimuli (5-ms duration) were produced at 6, 12, 18, 24, and 30 kHz using the SigGenRZ software and an RZ6 digital signal processor and delivered to the ear canal using a multi-field 1 (MF1) magnetic speaker (TDT). The stimulus intensity was increased from 10 to 90 dB SPL in 5-dB steps. ABR signals were fed into a low-impedance Medusa Biological Amplifier System (RA4LI, TDT), which delivered the signal to the RZ6 digital signal processing hardware. The recorded signals were filtered using a 0.5-1 kHz band-pass filter, and ABR waveforms in response to 256-tone bursts were averaged. ABR thresholds for each frequency were determined using the BioSigRZ software.
For DPOAE, a combination TDT microphone-speaker system was utilized. Primary stimulus tones were produced using an RZ6 digital signal processor with the SigGenRZ software and delivered using a custom probe with an ER 10B+ microphone (Etymotic, Elk Grove Village, IL) and MF1 speakers positioned in the ear canal. Primary tones were set at a frequency ratio with a frequency ratio (f2/f1) of 1.2 and equal intensity levels (L1 = L2). Intensity levels of the primary tones were increased from 20 to 80 dB SPL in 5-dB SPL increments. Fast Fourier transform (FFT) was performed at each primary tone for the DP grams and at each intensity for the I/O functions using BioSigRZ to determine the average spectra of the two primaries, the 2f1-f2 distortion products, and the noise floors.

Immunohistochemistry and histology
Immunoblotting and immunofluorescence were performed as previously described 55 were acquired by the maximum intensity projections of z-stacks for each segment by ImageJ.
Composite images showing the whole cochlea were constructed in Adobe Photoshop CS3 to display the entire turn of the cochlea. A frequency map of each specimen was drawn using ImageJ based on a previous report 57 . Phalloidin-and DAPI-positive IHCs and OHCs were counted in cochlear regions responsive to different sound frequencies, and segments containing dissection-related damage were omitted from the analysis. For SGN counting, fluorescence intensity of neurofilament heavy chains was quantified at the inner spiral plexus (spiraling underneath the IHCs) and unmyelinated outer spiral fibers (spiraling underneath the OHCs) using the auto-threshold algorithm ( "Otsu" method) 58 .

Cochlea live-imaging with thallium assay
The thallium flux assay was performed ex vivo to confirm potassium-induced physiological changes in the hair cells of AAV-injected and uninjected cochleae [59][60][61] . Briefly, 7-week-old mice with confirmed ABR-induced hearing improvement were euthanized with CO 2 to harvest both cochleae (i.e., AAV-injected and contralateral uninjected inner ears). The inner ear of age- wild-type cochleae treated with quinine (10 mM; an MET channel inhibitor) were used as negative control 63 . After loading the FluxOR dye, it was replaced with assay buffer (100 μL), and live-cell recording was performed using confocal microscopy and MetaFluor software.
When the baseline was stabilized, stimulus buffer (20 μL at a final concentration of 2 mM Tl + ) was added, and recording was performed for 120 s. After adding the stimulus buffer, △F120 was calculated relative to the baseline.

KCNQ4 mutation collection and gene editing efficiency prediction
For collecting all reported KCNQ4 mutations responsible for AHL in patients with deafness, we identified all KCNQ4 variants annotated as "pathogenic/likely pathogenic" in the ClinVar database and/or as "disease-causing mutations (DM)" in the HGMD database (professional v.2020.4). For a total of 49 pathogenic mutations in KCNQ4, combinations of sgRNAs and Cas9 variants were tested for gene editing efficiency in mutant alleles using our algorithm 50 . The best gene editing efficiency for each mutation was compared across all mutations and Cas9 variant types.

Statistical analysis
Data were pooled from at least three independent experiments and expressed as the mean ± SD.
Differences between two groups were determined using Student's t-test or two-way analysis of variance (ANOVA) with Bonferroni's corrections for multiple comparisons and conducted using GraphPad Prism v8.0 (GraphPad Software, San Diego, CA). Significant differences among the three groups were analyzed using one-way ANOVA with Tukey's multiple comparisons test for parametric comparisons and Kruskal-Wallis test with Dunn's multiple comparisons test for nonparametric comparisons. p < 0.05 was considered significant.

Data and software availability
The data and codes associated with this study are available from the corresponding author upon reasonable request. The published article includes all datasets/codes generated or analyzed during this study. Deep sequencing data have been deposited in the NCBI Sequence Read Archive (PRJNA691110).