RFX transcription factors are essential for hearing in mice

Sensorineural hearing loss is a common and currently irreversible disorder, because mammalian hair cells (HCs) do not regenerate and current stem cell and gene delivery protocols result only in immature HC-like cells. Importantly, although the transcriptional regulators of embryonic HC development have been described, little is known about the postnatal regulators of maturating HCs. Here we apply a cell type-specific functional genomic analysis to the transcriptomes of auditory and vestibular sensory epithelia from early postnatal mice. We identify RFX transcription factors as essential and evolutionarily conserved regulators of the HC-specific transcriptomes, and detect Rfx1,2,3,5 and 7 in the developing HCs. To understand the role of RFX in hearing, we generate Rfx1/3 conditional knockout mice. We show that these mice are deaf secondary to rapid loss of initially well-formed outer HCs. These data identify an essential role for RFX in hearing and survival of the terminally differentiating outer HCs.

S ensorineural hearing loss affects 1:500 newborns 1 and the majority of the elderly population 2 . The sensations of sound and movement are dependent on highly specialized post-mitotic mechanosensory cells called hair cells (HCs) 3 . Mammalian auditory HCs do not regenerate and their loss is a final common pathway in most forms of hearing dysfunction 4 . For this reason, understanding the molecular signalling cascades that lead to HC differentiation is important for hearing restoration. To date, several master regulators of HC fate and differentiation have been characterized. Among these are the transcription factors (TFs) ATOH1 (ref. 5), POU4F3 and GFI1 (refs 6-8). Nevertheless, forced expression of these three TFs in stem cells leads only to immature hair-cell-like cells 9 , underscoring the need to identify factors that mediate the differentiation and survival of maturating HCs. Furthermore, while the auditory and vestibular HCs and supporting cells (SCs) are structurally and functionally distinct, very few molecular differences between these cell types have been reported. Detailed knowledge of such markers, as well as regulators of terminal differentiation, is important to identify genes with a role in hearing and balance.
Gene expression analysis has been applied successfully to study development 10,11 , regeneration 12,13 , and identification of transcriptional cascades and molecular signalling pathways in the ear 14 . Given the complex structure of the inner ear sensory epithelia, cell type-specific analyses, either in the form of population analysis of sorted cells or in the form of single-cell analysis, have grown in favour [14][15][16] .
Here by performing a comprehensive cell type-specific comparison of the transcriptomes of HCs to other cell types from the auditory and vestibular systems of early postnatal mice, we identify the Regulatory Factor X (RFX) family of transcription factors as a key regulator of HC transcriptomes. Our results indicate an evolutionarily conserved role for RFX TFs in regulating the expression of genes encoding HC-enriched transcripts and demonstrate that RFXs are necessary for hearing in mice. In addition, we show that contrary to the known role of RFX as major regulators of cilia formation 17 , in RFX1/3 deficient HCs, the primary cilia (kinocilia) develop, and planar cell polarity is not impaired. The newly formed HCs seem structurally normal and functional until the outer HCs (OHCs) die rapidly at the onset of hearing, the time when the kinocilia are normally retracted. These data support a novel role for RFX in hearing, by maintaining the survival of normally formed HCs, probably through the regulation of their transcriptome during terminal differentiation.

Results
Inner ear cell type-specific gene clusters. To characterize the HC transcriptome in early post-natal auditory and vestibular systems, we used the Atoh1/nGFP transgenic mice expressing a green fluorescent protein (GFP) in all inner ear HCs 18 (Fig. 1a,b). Auditory and vestibular epithelia from inner ears of postnatal day 1 (P1) mice were separated into HCs, epithelial non-HCs (ENHCs) and non-epithelial cells (NECs) by flow cytometry (Fig. 1c, Supplementary Fig. 1). Gene expression levels were recorded from the sorted cells using whole genome expression microarrays (Supplementary Data 1). Hierarchical clustering, applied to all genes detected as expressed, showed a clear division of the samples based on cell types, namely HC, ENHC and NEC (Fig. 1d), demonstrating, as expected, that cell-type identity, rather than tissue of origin, is the major determinant of the cell transcriptome.
To define patterns of gene expression, we first searched for differentially expressed genes using an analysis of variance. We identified 6,556 probes, representing 4,269 unique genes (false discovery rateo5%) as differentially expressed between the cell types and tissues. Cluster analysis applied to this set of differentially expressed genes detected 12 main expression patterns ( Supplementary Fig. 2). The genes with a higher level of expression in HCs were divided into a cochlear-enriched cluster (cluster 1) and a vestibular-enriched cluster (cluster 3) (Fig. 1e). Functional enrichment analysis revealed that the cochlear HC cluster is significantly enriched for genes that regulate sensory perception of mechanical stimuli, whereas the vestibular HC cluster is significantly enriched for cytoskeleton and cilium-related genes (Fig. 1f).
In parallel, we used RNA-seq to record expression profiles of cochlear and vestibular HCs and ENHCs (Supplementary Data 2). There was a strong correlation between expression patterns measured by microarrays and RNA-seq, with RNA-seq measurements showing an improved dynamic range ( Supplementary Fig. 3a,b). Among the genes detected by RNAseq, the expression of 2,176 was elevated at least threefold in HCs compared with ENHCs in either the auditory or vestibular systems. In accordance with the patterns detected by microarray, the RNA-seq data set revealed two major expression clusters containing genes with increased expression in both auditory and vestibular HCs (clusters 1 and 2, Supplementary Fig. 4a,b). Moreover, the RNA-seq analysis detected not only HC-enriched protein-coding transcripts but also a multitude of long, noncoding transcripts whose expression is elevated in HCs (Supplementary Data 2).
To determine whether differential expression can be predictive of functional significance, we examined the expression of mouse orthologues of the known human deafness-causing genes detected in the microarray and RNA-seq data sets. Interestingly, out of 70 orthologues of known human deafness-causing genes, 56 and 64 were detected as expressed in at least one cell type, in the microarray and RNA-seq, respectively. Of these, in the microarray data set, 41 genes were differentially expressed Post-sort analyses of sorted cells showing a cell type-specific purity 494% for HCs, 499% for NECs and 498% for ENHCs. (d) Hierarchical clustering, based on all the expressed genes in the data set, separated the probed samples into three major branches according to cell type: HC, NEC and ENHC (c-cochlea and v-vestibular samples). (e) Two major clusters containing genes whose expression is elevated in HCs. Each cluster is represented by the mean pattern of its genes' expression levels (error bars: ±s.d. calculated over all genes assigned to each cluster). Expression levels of each gene were standardized (mean ¼ 0, s.d. ¼ 1) prior to clustering. This methodology enabled grouping of genes that share similar patterns, but not necessarily the same magnitude of expression. The entire set of clusters is shown in Supplementary Fig. 2. (f) Enriched GO categories with a P value o0.01 (hypergeometric test) after false discovery rate correction for multiple testing. (g) Marker genes specific for cochlear HCs (top) and vestibular HCs (bottom) as detected in the expression microarray data set (left) and confirmed by the RNA-seq data set (right). Red represents enrichment, Green represents depletion. Data represent results from three biological replicates. ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/ncomms9549 (enrichment P value ¼ 1.6E À 9; hypergeometric test), and statistically over-represented in the clusters of genes whose expression is elevated in HCs (29 genes, P value ¼ 5.5E À 6; hypergeometric test). Similarly, in the RNA-seq, 33 genes showed at least threefold elevated expression in HCs compared with ENHCs in either the auditory or vestibular systems (P value ¼ 4.2 E À 11; hypergeometric test; Supplementary Data 3; and see examples in Supplementary Fig. 4c). With B100 deafness loci remaining to be cloned, these data suggest that differential inner ear expression and specifically elevated expression in HCs, could be used as a valuable criterion in prioritizing genes/variants for validation.   ARTICLE Identification of inner ear cell type-specific markers. Next, to further define cell type-specific marker genes within each tissue (namely genes whose expression is strictly confined to a particular cell-type in our examined panel), we applied stringent expression criteria to the microarray data set (Methods). This analysis identified 158 annotated genes as markers of inner-ear HCs (out of 276 probes corresponding to annotated and predicted genes; Supplementary Data 4). Most of these genes marked both the cochlear and vestibular HCs, and included many genes that, in the newborn mouse inner ear, are known markers of HCs, such as Atoh1, Pou4f3 and Gfi1. Egfp was identified as a HC-specific marker validating the specificity of the sorting approach. Of interest, 11 genes were specific for the cochlear HCs while 13 were specific for the vestibular HCs (Fig. 1g). These included Slc26a5, a known cochlear HC-specific marker (uniquely expressed in OHCs 19 ), and Ocm which in the newborn inner ear is not detected in cochlear HCs but is already expressed in vestibular HCs 20 . Similarly, we identified 32 unique genes as markers of inner ear ENHCs of which 7 were specific for the cochlea and 4 were specific for the vestibular system (Supplementary Data 4).
To validate the identified cell type-specific markers, we examined our RNA-seq measurements and found a remarkable confirmation for the specificity of our marker sets ( Fig. 1g; Supplementary Fig. 5). Specifically, of the 158 genes defined by the microarray as markers of HCs, 155 showed in the RNA-seq data at least 2-fold enrichment in HCs, with an average of 10-fold enrichment in cochlear HCs and 14-fold enrichment in vestibular HCs (Supplementary Data 4). As for the set of ENHC markers, for all except 1 gene, the expression levels measured by RNA-seq were at least 2-fold enriched in ENHCs compared with HCs (with an average of 11.3-fold and 7.5-fold for the cochlear and vestibular ENHCs, respectively). The stringent criterion we set for calling genes as HC-markers favours specificity (low rate of false positives) and over sensitivity (low rate of false negatives), and accordingly many genes whose expression is enriched in HCs were not called as marker genes ( Supplementary Fig. 5). Overall, these data define over 150 validated HC and ENHC cell type-specific and tissue-specific markers in the newborn mouse inner ear.
RFX are important regulators of the HC transcriptome. Our next goal was to identify transcription factors that regulate the expression of genes encoding HC-enriched transcripts. We and others have demonstrated that computational analysis of cisregulatory elements in promoters of differentially expressed genes is a powerful tool for the identification of transcriptional regulators 10,21,22 . This reverse-engineering approach for delineating transcriptional networks is based on the assumption that genes co-expressed in the same cell-type(s) are likely to be regulated by common transcription factors that are specifically active in those cells. Thus, these genes are expected to share specific DNA regulatory elements in their promoter regions. Therefore, we computationally searched for DNA motifs whose prevalence in the promoters of genes encoding HC-enriched transcripts, as detected by our microarray and RNA-seq data, is significantly higher than in the promoters of all other genes in our data set. We defined promoters as the genomic region from 1,000 nucleotides (nt) upstream to 500 nt downstream of the transcription start site (TSS). A single, highly statistically significant DNA motif was identified as over-represented in promoters of the genes encoding HC-enriched transcripts. This motif corresponds to the binding signature of the RFX family of transcription factors (Fig. 2a), and its over-representation was most prominent on promoters of genes encoding HC-enriched transcripts, which showed higher elevated expression in vestibular compared with auditory HCs (cluster 3 in Fig. 1e and cluster 1 in Supplementary Fig. 4a). The RFX signature was detected on promoters of 225 genes encoding HC-enriched transcripts, resulting in a B3-fold over-representation of the motif (P value ¼ 1.5E À 37; hypergeometric test) and tagging the corresponding genes as putative direct targets of RFX in HCs (Supplementary Data 5). A similar sequence analysis using promoters of the human and zebrafish orthologous genes also pinpointed the RFX motif as highly enriched and the top scoring motif in both organisms, suggestive of an evolutionarily conserved pathway (enrichment P value ¼ 2.9E À 30 and 9.2E À 29 for human and zebrafish promoters, respectively; hypergeometric test; Supplementary  Fig. 6). Notably, the putative targets include Lrrc51, a known deafness-causing gene 23 29 , the 225 putative RFX target genes were enriched for cytoskeleton and cilium genes (Fig. 2b,c). Overall, the expression of the putative RFX targets was significantly enriched in both the cochlear and the vestibular HCs, with the vestibular showing a greater level of elevated expression ( Supplementary Fig. 7).
We next analysed two recent expression data sets that also recorded gene expression in the inner-ear and defined sets of HC-enriched genes. One of these studies purified cochlear HCs using the Atoh1 A1GFP/A1GFP mouse 30 , and the other used Pou4f3-GFP mice to isolate auditory and vestibular HCs for expression analysis 31 . Notably, in both of these data sets we detect the RFX binding signature as the most highly over-represented motif in the promoters of genes encoding for HC-enriched transcripts ( Supplementary Fig. 8a). We also find that the expression of the set of putative RFX target genes derived from our combined microarray and RNA-seq data sets is highly enriched in HCs as measured by these two independent studies ( Supplementary Fig. 8b,c).
RFX targets identified by ChIP-seq are elevated in HCs. To validate and further identify which of the candidate target genes are bound in vivo by RFX, we analysed a chromatin-immunoprecipitation sequencing (ChIP-seq) data set performed with antibodies for RFX1 and RFX3 in the mouse Min6 cell line 32 . In this data set, RFX1 and RFX3 were found to bind the promoters of 2,540 and 3,575 genes, respectively, with a highly significant overlap ( Supplementary Fig. 9a). We found that as a group, the expression of ChIP-seq-identified RFX target genes was significantly elevated in HCs compared with their expression in ENHCs and NECs ( Fig. 2d; Supplementary Fig. 9b). This result further supports the computational prediction of a role for RFX in regulation of gene expression in HCs. Moreover, intersection between the computationally predicted RFX targets in HCs and the ChIP-seq data confirmed 142 of the 225 candidate promoters as bound in vivo by at least 1 of the probed RFX TFs and 130 of these promoters were bound by both RFX1 and RFX3, indicating a high degree of overlap in DNA binding between RFX members in the HCs (Supplementary Data 5).
The role of RFX in HCs is evolutionarily conserved. We extended our analysis to another species, to determine whether the role of RFX in regulating the HC transcriptome is evolutionarily conserved, and utilized the ppv3b:GFP transgenic zebrafish, which express GFP predominantly in HCs 33 (Fig. 3a,b). We sorted GFP-positive ( þ ) and negative ( À ) cells from 5-day post fertilization (dpf) larvae using flow cytometry (Fig. 3c) and profiled their transcriptomes using RNA-seq (Supplementary Data 6). We observed a 3-to 37-fold enrichment for known HCenriched transcripts in the GFP( þ ) cells, confirming their identity as HCs (Fig. 3d). Consistent with an evolutionarily conserved role for RFX in regulating the HC transcriptome, expression of the zebrafish orthologues of the 225 putative RFX target genes identified in mouse was significantly elevated in the zebrafish GFP( þ ) cells (Fig. 3e). Conversely, zebrafish orthologues of mouse inner ear vascular endothelial-or neuronalenriched transcripts were not elevated in the zebrafish GFP( þ ) cells, validating the specificity of our result ( Supplementary  Fig. 10a,c).
As a means of independent corroboration, we analysed a published data set of zebrafish HCs compared with SC gene expression, generated from 5 dpf alpl:mCherry;pou4f3:GFP larvae 34 . Again, the RFX motif was the only statistically overrepresented signature detected in the promoters of genes with enriched expression in zebrafish HCs ( Supplementary Fig. 10d). In this data set too, expression of the zebrafish orthologues of our 225 putative RFX targets was significantly elevated in HCs (Fig. 3f).
RFX transcription factors are essential for hearing. To determine whether RFX are necessary for hearing, we first queried our RNA-seq data and found that Rfx3 is enriched in HCs and the most abundant RFX transcript in the inner ear (Supplementary (a) De-novo motif analysis revealed that the promoters of genes whose expression is elevated in inner ear HCs are significantly enriched for a DNA motif that matches the binding signature of the RFX family of TFs (the known RFX motif is from TRANSFAC DB 70 ; accession number M00281). The enriched motif was detected in promoters of 225 HC-elevated genes (enrichment factor ¼ 2.7; P value ¼ 1.5E À 37). (b) GO functional categories that were over-represented in the set of 225 putative RFX HC target genes. (c) Expression pattern of selected putative RFX targets that function in ciliogenesis (n ¼ 3; error bars: ± s.d.). (d) HC enrichment was defined per gene as the ratio between its expression in HCs and its average expression over the three probed cell types (HC, ENHC and NEC). The plot compares the cumulative distribution of these HC-enrichment measures (in log2 scale; x-axis) between the set of all RFX1 targets defined by ChIP-seq analysis (red) and all the other genes in the data set (blue). The distribution for the RFX1 targets is significantly shifted to the right, demonstrating that their expression level as a group is markedly elevated in HCs (P value ¼ 1.3E À 77; Wilcoxon test). See Supplementary Fig. 9b for similar analysis applied to ChIP-Seq targets of RFX3. NATURE COMMUNICATIONS | DOI: 10.1038/ncomms9549 ARTICLE Data 2). In addition, immunostaining detected a robust nuclear expression of RFX3 in all HCs of the auditory and vestibular systems along with much weaker expression in ENHCs (Fig. 4, Supplementary Fig. 11). Therefore, RFX3 is likely to be an important TF-regulating HC development. To determine the role of RFX proteins in HCs, we generated conditional knockout (cKO) mice. First, we used the Gfi1-Cre mice to delete Rfx3 from all inner ear HCs. Hearing measurements of 3-month-old Rfx3 cKO revealed normal thresholds which could, however, represent functional redundancy with other RFX family members (Fig. 5). We next examined which additional RFX family members are expressed in HCs and detected expression of Rfx1, Rfx2, Rfx5 and Rfx7. We generated cKO for Rfx1 ( Supplementary Fig. 12), because RFX1 and RFX3 share conserved structural domains 35 , show a high degree of overlap in their DNA binding to candidate genes, and form heterodimers in many cell types 36,37 . As with Rfx3 cKO, we observed normal hearing in the Rfx1 cKO (Fig. 5). However, remarkably, the Rfx1/3 double cKO mice suffered from rapidly progressive hearing loss with significantly elevated hearing thresholds already at P22 progressing to no measurable hearing by 3 months of age (Fig. 5).
Rapid OHC loss in Rfx1/3 cKO mice. To determine the cause of the hearing loss in the Rfx1/3 double cKO mice, we performed whole mount phalloidin staining and Scanning Electron Microscopy (SEM). In the mutant mice, the sensory epithelium appeared normal at P12 throughout the length of the cochlear duct, with no significant loss of HCs (Fig. 6a). In contrast to P12, there was a widespread loss of outer but not inner HCs by P15, consistent with the elevated hearing thresholds at P22 (Fig. 5a). In addition, in contrast to wild-type inner HCs (IHCs), Rfx1/3 cKO IHCs had retained kinocilia at P15 (Fig. 6b), suggesting that RFX1/3 are involved in normal retraction of this organelle in the late development, at least in IHCs. By P90, we observed near complete loss of all OHCs throughout the length of the cochlear -5 (e) Expression of the zebrafish orthologous genes of the 225 putative targets of RFX that we identified in murine HCs, was significantly elevated in zebrafish HC (GFP( þ ) cells). Ratios between expression level in the GFP( þ ) and GFP( À ) cells were calculated for each gene in the zebrafish data set. The plot compares the distribution of these ratios (in log2 scale; y-axis) between the set of orthologues of the RFX targets (left boxplot) and all the other genes in the data set (right boxplot). The box indicates the first and third quartiles; the horizontal band inside the box indicates the median. The whiskers extend to the most extreme data point, which is no more than 1.5 times the interquartile range from the box. P value calculated using Wilcoxon test. (f) Similar to panel (e), but using an independent zebrafish gene expression data set. Here too, expression of RFX target genes is significantly elevated in HCs as compared with their expression levels in non-HCs. duct and progressive IHC degeneration associated with sporadic fusion of stereocilia on some of the remaining IHC stereociliary bundles (Fig. 6c).
To further clarify the mechanism of the loss of OHCs, we analysed the expression of select cell death markers by western blot analysis. We detected a decrease in the expression of BCL2 (pro-survival factor), along with an increase in caspase-3 and BAX (pro-apoptotic factors), in the P12 and P15 mutant cochlea compared with P15 controls (Fig. 6d, Supplementary Fig. 13). Thus, despite the intact appearance of the sensory epithelium, cell death has already initiated at P12, leading to a rapid and diffuse loss of OHCs by P15. These data indicate that a lack of RFX expression leads to programmed cell death in the OHCs.
Early differentiation of OHCs in Rfx1/3 cKO mice is normal. As we observed an accelerated degeneration of HCs at P15, we next performed structural and functional studies to examine the HCs in the earlier postnatal period. We first evaluated the HC polarity using confocal microscopy, because the RFX TFs are associated with ciliogenesis 17 . Abnormalities of ciliogenesis in the inner ear usually result in the disruption of planar cell polarity 38 . We found that Rfx1/3 cKO stereocilia bundles were properly polarized throughout the length of the cochlear duct (Fig. 7). We further evaluated the structure of the hair bundles using SEM and found that Rfx1/3 cKO HCs at P8 have well-formed bundles, including intact primary cilia (kinocilia), stereocilia, tip-links and inter-stereociliary links (Fig. 8a,c). To determine whether OHCs express markers associated with functional maturation in the first postnatal week, we stained cochleae for the OHC-specific molecular motor prestin 19 , and found an identical expression of prestin in the mutant and wild-type mice (Fig. 8d). Finally, consistent with functional mechanotransduction 39 , the newborn HCs of the Rfx1/3 cKO exhibited intact uptake of FM1-43 (Fig. 8e). These data demonstrate that early development and  ARTICLE maturation of HCs is not affected in the Rfx1/3 cKO, and identify a critical role for RFX1/3 proteins in regulating the survival of the post-natal terminally differentiating OHCs.

Discussion
In this study, we employed a genomic cell type-specific approach to comprehensively profile the HC transcriptome. Our data revealed numerous HC and ENHC marker genes, common and divergent between the auditory and vestibular systems at P1. We identified and validated 183 HC and ENHC cell type-specific markers, many of which have not been previously described in the ear. Several of these markers include genes encoding regulatory proteins (for example, Insm1, Pou2af1), ion channels (for example, Ano4), kinases (for example, Fn3k, Nek5) or ubiquitin ligases (for example, Trim36). The list also includes numerous genes with a known function in inner ear development.
In addition, using an analysis of expression levels of genes encoding for homologues of known deafness-causing genes, we found that differential cell type-specific expression in the ear can be suggestive of functional significance. Taken together, and given the high stringent criteria used to identify these newborn inner ear cell type-specific markers, we anticipate that many of their encoding genes carry critical roles in inner ear development. Gene expression levels are often not sufficient to identify key regulatory pathways, because numerous transcription factors are expressed in any given cell type, and the level of expression does not necessarily correlate with a functional significance. Using an unbiased reverse-engineering approach for de novo identification of transcriptional regulators of gene expression, we identified the family of RFX proteins as evolutionary conserved regulators of the HC transcriptome. Furthermore, analysis of the single cell RNA-seq data in the accompanying study by Burns et al. 40 identifies RFX binding motifs as the most enriched in the set of genes whose expression increases in differentiating HCs, consistent with our results. We also detected the RFX-binding  Based on the role of RFX in ciliogenesis, we anticipated either the loss of kinocilia in the mutant HCs or functional defects associated with ciliogenesis-such as planar cell polarity defects 38 . The normal appearance of the HCs in the young postnatal Rfx1/3 cKO mice was therefore surprising. In addition, the expression of the 225 predicted RFX target genes in the vestibular HCs at P1 is higher than the expression of these genes in the auditory HCs ( Supplementary Fig. 7). Therefore, the normal behavioural vestibular phenotype in Rfx1/3 cKO mice was also surprising. However, we cannot exclude possible compensatory function of other RFX genes expressed in vestibular HCs, such as Rfx2, Rfx5 or Rfx7. The absence of abnormalities of ciliogenesis or planar cell polarity in the mutant mice, coupled with the timing of HC death, only after P12, suggest a novel role for RFX TFs in hearing, independent of ciliogenesis. The exact mechanism of OHC degeneration in Rfx1/3 cKO mice is yet unclear, but the loss of outer but not inner HCs could result from the well-known greater susceptibility OHCs to cochlear insults or from compensatory effects of other Rfx genes. In addition, differential requirement of TF for HC survival has been previously described; For example, while Barhl1 is expressed in both auditory and vestibular HCs, it is only required for the survival of cochlear HCs 41 . Likewise, while Gfi1 is expressed in all cochlear HCs, OHCs show a greater dependence on its expression compared with IHCs 7,42 .
Our data show that Rfx1/3 are necessary for OHC survival at the latest stages of their development. In the absence of Rfx1/3, OHCs undergo very fast programmed cell death. The timing of the OHC death in the Rfx1/3 cKO mice coincides with two critical processes both related to RFX biology. First, the kinocilia, which are normally resorbed in wild-type HCs during this period, are retained in the mutant mice. This suggests a largely unappreciated but possible link between normal recess of kinocilia and HC survival. Second, the endocochlear potential is also developed exactly at this developmental stage 43 . The appearance of endocochlear potential should dramatically increase the resting current through the mechanotransduction channels, Ca 2 þ influx in the HCs through these channels and associated metabolic stress in the OHCs 44 . Metabolic stress after noise exposure has been shown to result in DNA damage 45 . Interestingly, RFXs have a known role in cellular homeostasis through regulation of DNA integrity 46,47 . Therefore, it is possible that RFX may regulate HC survival also through maintenance of DNA integrity.
Perhaps the most significant contribution of this study is in the identification of RFX as a new family of TFs with a role in regulating the HC transcriptomes. Previously described transcription factors important for HC differentiation such as ATOH1 (ref. 5), POU4F3 (refs 8,42,48) and GFI1 (refs 7,42) are necessary for early (embryonic) HC differentiation and survival, while RFX1/3 are crucial for the latest stage of HC development. In Drosophila, rfx is directly regulated by atonal (the Atoh1 homologue), and in conjunction with a forkhead transcription factor, fd3F, regulates the differentiation of the chordotonal (mechanosensory) neurons involved in hearing in Drosophila 49,50 . In mouse, the closest orthologue to fd3F is Foxj1, which is also expressed in the auditory and vestibular HCs (Supplementary Data 2), and could be required together with RFX to regulate the maturation of HCs. Interestingly, a recent manuscript by Costa et al. 9 showed that forced expression of ATOH1, POU4F3 and GFI1 in stem cells results in immature HC like-cells, which specifically lack the expression of the orthologues of some of the target genes of fd3F, consistent with our hypothesis   Generation of conditional KO mice. Generation of Rfx3 flox/flox mice is described in Bonnafe et al. 53 . Rfx1 flox/flox were generated as follows: the targeting vector contained two loxP sites (P) inserted into the introns flanking exon 10 of the Rfx1 gene, and an frt (F) flanked neomycin resistance (pgk-neo-pA) cassette in the intron downstream of exon 10 ( Supplementary Fig. 12). Exon 10 contains an essential portion of the Rfx1 DNA-binding domain. The linearized targeting vector was transfected into embryonic stem (ES) cells by electroporation and G418resistant clones were selected. Screening for homologous recombination was done by PCR using an internal primer and a primer situated upstream of the 5 0 end of the targeting construct. Positive clones were confirmed by Southern blot analysis  using 5 0 , 3 0 and internal probes. The neo resistance gene was excised in recombinant ES cell clones by transfection with the pCAGGS-Flpe vector and selection with puromycin. Clones were isolated and screened for G418 sensitivity. Excision of the neo resistance gene was controlled by Southern blot analysis with an internal probe. Information on primers, restriction sites and probes will be made available on request. Correctly targeted ES cell clones were injected into C57BL/6J blastocysts and implanted into recipient females. Germ line transmission of the recombinant Rfx1 allele was confirmed by PCR and Southern blotting. Conditional deletion of exon 10 was achieved by breeding in a Gfi1-Cre transgene. After deletion of exon 10, splicing of exon 9 to exon 11 leads to a frame shift. For generation of mice with a conditional deletion of Rfx1 or Rfx3, Rfx1 flox/flox and Rfx3 flox/flox mice were crossed with Gfi1-Cre knock-in mice 54  Cell separation by flow cytometry. Newborn (P1) Atoh1/nGFP mice were used to separate the inner ear cells to HCs, ENHCs and NECs. Inner ears from four to eight mice were used for each biological replicate. Tissue harvesting and processing of mouse inner ears for flow cytometry was performed as previously described without modifications 10 . The auditory sensory epithelia and vestibular tissues were collected in 48-well plates containing 0.5 mg ml À 1 thermolysin (Sigma) and incubated for 20 min in a 37°C/5%CO2 humidified tissue culture incubator to allow partial digestion of the extracellular matrix. The thermolysin was then aspirated, replaced with Accutase enzyme cell detachment medium (eBioscience) and tissues were incubated for 3 min in a tissue culture incubator followed by mechanical disruption using a 23G blunt ended needle connected to a 1 ml syringe. The 3-min incubation and mechanical disruption were performed twice. The reaction was stopped by adding an equal volume of IMDM supplemented with 10% heat-inactivated foetal bovine serum (FBS). After confirming cellular dissociation by direct visualization using an inverted tissue culture microscope, cells were passed through a 40-mm cell strainer (BD) to remove cell clumps. Following dissociation, we stained the cells with CD326-APC (1:2,000, Biolegend) diluted in IMDM-10% FBS for 30 min on ice. CD326 is a marker for the epithelial cells in the inner ear 10 . Cells that were positive both for GFP and CD326 are HCs, positive for CD326 and negative for GFP are ENHCs, and negative both for GFP and CD326 are NECs. Of note, possible minor expression of GFP in neuronal cells 55 was eliminated from the sorted cells by restricting HCs to cells that are both GFP and CD326 positive. Pillar cell expression of GFP was not strongly visualized in our prep as demonstrated in Fig. 1a. For cell separation from ppv3b:GFP zebrafish, embryos were dissociated as follows (previously reported 56 ): Five dpf larvae were dissociated by incubation in trypsin (Trypsin-EDTA 0.25%, Life Technologies) for 30 min with intermittent mechanical disruption by trituration. The reaction was stopped by adding HBSS supplemented with 10% FBS and 100 mg ml À 1 DNaseI. Cells were filtered through a 70-mm cell strainer (BD Biosciences), centrifuged at 800g for 7 min at 4°C and resuspended in HBSS supplemented with 10% serum. Following each sort, a small aliquot of cells was re-analysed to determine cell purity. Gene expression profiling and analysis. Microarrays data analysis. Expression profiles were recorded from six cellular populations (three cell-types: HCs, ENHCs and NECs, isolated from two inner ear organs: the cochlea and the vestibular system). These six conditions were probed in biologically independent triplicates using Illumina MouseRef-6 v2.0 Expression BeadChips (Illumina, San Diego, CA, USA). These arrays contain 445,000 probes which collectively interrogate all RefSeq annotated mouse genes (426,000 genes) and B7,000 RefSeq predicted genes. Expression levels were calculated using Illumina's Bead-Studio package.
Probes not readily detected in the data set were filtered out using detection P values assigned by Bead-Studio to each measurement and requiring that each probe be detected (P valueo0.01) in at least two samples. This criterion left 23,051 probes, corresponding to 17,275 annotated and predicted genes for subsequent analyses. Arrays were then normalized using quantile normalization. Hierarchical clustering, using average-linkage method, was applied to assess overall similarity between the different conditions. Differentially expressed genes in the data set were identified using two-way analysis of variance (ANOVA) analysis. 6,556 probes were significantly (false discovery rate o5%) associated with either factor (cell-type or organ) or showed significant interaction between these two factors (that is, different cell-type expression patterns between the two organs). Main expression patterns exhibited by the set of differentially expressed probes were delineated by cluster analysis using the CLICK algorithm implemented in the EXPANDER package 57 . Prior to clustering, expression levels of each probe were standardized to mean ¼ 0 and s.d. ¼ 1. To define cell type-specific marker genes, we first examined the microarray data set for cutoff values that correspond with 'expression' compared with 'no-expression' of probes, and observed that that all probes flagged as not expressed (that is, detection P value40.05) had expression levels o120 RNA-seq data analysis. Sequenced reads were aligned to the mouse and zebrafish genomes (mm9 and Zv9) using TopHat 58 . Based on profiles of base-calls quality, the reads in the mouse data set were trimmed and only the first 50 nt were used for the alignment. Gene expression levels (read counts) were calculated using HTseq 59 based on ENSEMBL gene annotations 60 . The mouse and zebrafish data sets were then normalized using quantile normalization and then a floor value of 20 reads was set (to avoid inflation of fold-change estimates due to division by low counts). In the mouse and zebrafish data sets, 12,814 and 17,668 genes were detected as expressed, respectively.
GO analysis. Enriched Gene Ontology (GO) categories in the clusters were detected by DAVID 61 using the set of all genes expressed in the data set as the background set.
Analysis of cis-regulatory motifs. Over-represented sequence motifs in promoters were searched using AMADEUS 62 . The set of promoters of all genes detected as expressed in each analysed data set was used as the background set. The promoter regions examined spanned from 1,000 nt upstream to 500 nt downstream the TSS. Mouse, human and zebrafish orthologous genes were mapped using ENSMART utility 60 . In all tests of HC-elevated genes, the top scoring motif matched the known binding signature of RFX TFs. The location distribution of this motif consistently showed a very sharp peak at the TSS ( Supplementary Fig. 6b), indicating that RFX commonly regulates the expression of its target genes by binding to their very proximal promoter region. Accordingly, the enrichment was most statistically significant when the search was narrowed to the interval between 200 nt upstream and 100 nt downstream of the TSSs. The results reported correspond to this interval.
Deafness-related loci. A list of cloned human deafness-related genes and deafness loci for which the underlying gene was not cloned yet was compiled from the Hereditary Hearing loss Homepage 63 . Human genes located within deafness loci were extracted using a Perl script. Mouse homologues of these genes were found using NCBI's HomoloGene 64 .
ChIP-seq. ChIP was performed as previously described 65 : MIN6 beta cells were crosslinked with 1% formaldehyde for 10 min at room temperature. The reaction was stopped by adding of glycine to 0.125 M followed by cold PBS washes. Cells were lysed in TE buffer supplemented with protease inhibitors and 0.5% NP-40. Following centrifugation, the nuclei were lysed in TE supplemented with 0.5 M NATURE COMMUNICATIONS | DOI: 10.1038/ncomms9549 ARTICLE NaCl, 1% Triton X-100, 0.5% sodium deoxycholate and 0.5% sarcosyl. Crosslinked chromatin was resuspended in TE plus 100 mM NaCl, sheared by sonication and cleared by two cycles of centrifugation at 15,000g for 15 min. About 10 mg of the resulting sheared chromatin was diluted 10-fold with IP buffer (20 mM Tris-HCl, pH 8.0, 200 mM NaCl, 2 mM EDTA, 1 mM phenyl-methyl-sulfonyl-fluoride, 0.1% sodium deoxycholate, 0.1% SDS) containing 50 mg ml À 1 salmon sperm DNA, 100 mg ml À 1 Escherichia coli tRNA and 1 mg ml À 1 bovine serum albumin. Diluted chromatin was then incubated at room temperature with protein A-Sepharose beads, polyclonal anti-RFX1 and anti-RFX3 sera 37 . Beads were washed twice with IP buffer, twice with IP buffer containing 500 mM NaCl, twice with 20 mM Tris-HCl, pH 8.0, 0.25 M LiCl, 2 mM EDTA, 0.5% NP-40, 0.5% sodium deoxycholate and once with TE containing 0.1% NP-40. Chromatin fragments bound to the beads were eluted by incubation in 100 mM Tris-HCl, pH 8.0, 1% SDS for 10 min at 65°C. Eluted chromatin was digested for 1 h at 37°C with 400 mM NaCl, 200 mg ml À 1 proteinase K, followed by incubation at 65°C overnight to reverse the crosslinks. DNA was extracted with phenol-chloroform, resuspended in TE and quantified with a Qubit Fluorometer from Invitrogen. To generate the ChIPsequencing library, 10 ng of DNA were used as input for the ChIP-seq Sample Preparation Kit (Illumina). Sequencing was performed with the Genome Analyser II (Illumina) using the 36 Cycle Sequencing Kit v2. Sequences were derived from a single channel of the flow cell. Data were processed using the Illumina Pipeline Software package v1.0.
Immunohistochemistry on paraffin sections. Immunohistochemistry reactions were performed on 5-mm sections of paraffin-embedded inner ears from P1 wildtype ICR mice. Sections were deparaffinized with xylene and rehydrated with a decreasing ethanol gradient, followed by antigen retrieval with sodium citrate buffer (pH6). Sections were blocked with PBS supplemented with 0.2% Tween-20, 10% normal goat serum and 1% bovine serum albumin for 30 min at room temperature before incubation with rabbit polyclonal anti-RFX3 (HPA035689, Sigma-Aldrich, 1:200) overnight at 4°C. Antibody labelling was performed with a biotin-conjugated anti-rabbit IgG (1:500, Life Technologies) and detection with Alexa Fluor 594-conjugated streptavidin (1:500, Life Technologies) for 1 h at room temperature. The nuclei were counterstained with 300 nM of DAPI (Life Technologies). Samples were then mounted in anti-fade medium (ProLong Gold antifade reagent, Life Technologies). Images were acquired with a Nikon Eclipse E800 coupled with a Qimaging Retiga EXi Fast1394 and analysed with the Volocity Image Analysis Software (Perkin Elmer).
Immunohistochemistry on whole mounted samples. Sensory epithelia from cochlear and vestibular systems were dissected from mice at the ages specified in the figure legends. Tissues were permeabilized in PBS-0.5% tween20, blocked in M.O.M. Mouse Ig Blocking Reagent (Vector Laboratories) for 1 h at room temperature and incubated with mouse monoclonal Anti-Acetylated Tubulin (Sigma) diluted in M.O.M. Diluent at 1:500 overnight at 4°C or blocked in PBS-0.2% tween20-10% Donkey serum and incubated with a goat anti-Prestin antibody (1:200, SC22692, Santa Cruz Biotechnology). Detection was performed with an Alexa Fluor 546 conjugated anti-mouse IgG (1:500, Life Technologies) or a Alexa Fluor 546 conjugated anti-goat IgG (1:500, Life Technologies) supplemented with Alexa Fluor 488 conjugated Phalloidin (1:500, Life Technologies) and 300 nM of DAPI for 1 h at room temperature. Samples were then mounted in anti-fade medium (ProLong Gold antifade reagent). Images were acquired with a Zeiss LSM 510 confocal microscope. For quantification of planar cell polarity, sensory epithelia from 3-day-old mice were obtained and labelled for acetylated tubulin as described, the orientation of 20 cells from the middle and basal turns of each cochlea (consisting of five HCs from each row) was measured using the ImageJ software (http://imagej.nih.gov/ij/) as follows: a line perpendicular to the row of pillar cells and extending through the HCs was drawn, as well as a line going from the centre of each HC to their kinocilium. The orientation of the bundle is calculated by measuring the angle obtained at the intersection of these two lines. Graphs were plotted using GraphPad Prism and statistical analyses were performed using a one way ANOVA, post test: Tukey's test.
FM1-43 staining. Sensory epithelia from cochleae were dissected from P1 and P3 mice and cultured for 24 h on CellTak-treated glass bottom culture dish in DMEM/ F12 (Life Technologies) supplemented with 10% FBS and 10 mg ml À 1 ampicillin. Culture dishes with adherent cochlear explants were rinsed once with HBSS supplemented with 10 mM HEPES pH 7.2 (HHBSS), followed by incubation with 3 mM FM1-43 (Life Technologies) for 10 s, immediately followed by four washes with HHBSS. Images were acquired with a Zeiss LSM 510 confocal microscope.
Scanning electron microscopy. The temporal bones from 8-, 15-, 22-and 90-dayold mice were dissected in ice-cold Leibowitz L-15. Cochleae were exposed and small holes were made at the apex and the base to allow fixative penetration. Cochleae were fixed in 2.5% glutaraldehyde in 0.1 M cacodylate buffer supplemented with 2 mM CaCl 2 for 1.5 h at room temperature or at 4°C overnight. Then, the cochlear partition was dissected in an ice-cold PBS or distilled water. Dissected specimens were dehydrated in a series of ethanol, followed by critical point drying from liquid CO 2 , and sputter coated with platinum (5.0 nm, controlled by a film-thickness monitor) and observed with a field-emission electron column of Helios 660 NanoLab instrument (FEI Co., Hillsboro, OR). Alternatively, a modified osmium-thiocarbohydrazide-osmium procedure was used. Briefly, the fixed and dissected specimens were treated with 1% osmium tetroxide three times with a 30min water wash and a 20-min incubation in saturated thiocarbohydrazide between each osmium tetroxide treatment. Then, the specimens were dehydrated as described above and imaged with Quanta 200 SEM (FEI).
Auditory brainstem response. Tone burst stimuli at 8, 16 and 32 kHz were presented via high frequency transducers coupled at the ear via polyethylene tubing tubing. Subcutaneous electrodes were placed at the nuchal crest, posterior to the left pinna and at the hip for noninverting, inverting and ground leads. Traditional signal averaging was used where ongoing electroencephalographic activity was amplified (200,000 Â ), filtered (300 to 3,000 Hz), digitized (100 kHz) and averaged (256 to 512 averages per waveform). ABR intensity series was collected with a descending series of stimulus levels (5-dB steps) beginning at 110 dB peSPL. ABR thresholds were determined and compared between knockout and controls using ANOVA.
DPOAE. DPOAE frequencies (f1, f2, f2/f1 ratio ¼ 1.25) at equal levels (L1 ¼ L2 ¼ 60 dBSPL), 150 ms duration, were generated and routed through separate drivers to mix acoustically in the ear canal. Geometric mean ((f1 Â f2) 0.5 ) frequencies ranged from 6.0 to 48.5 kHz (eight frequencies per octave). Ear canal sound pressure levels were recorded with a calibrated low noise probe microphone. Microphone output was amplified, sampled (200 kHz) and subjected to fast Fourier transform (FFT). The amplitude of f1, f2 and the 2f1-f2 distortion products were measured from the FFT waveform. The noise floor was measured from the amplitudes in the five frequency bins above and below the 2f1-f2 component. The mean DPOAE amplitude across all tested primary frequency pairs was calculated and compared between knockouts and wild-type controls using ANOVA.
Western blot analysis. Cochlear ducts were dissected from P12 and P15 wild-type and Rfx1/3 cKO mice and flash frozen in dry ice. Each biological replicate consisted of tissue collected from four ears. Two biological replicates (a total of eight ears) were collected for analysis from each time point and genotype. Cochlear tissues were subjected to lysis in RIPA buffer (Pierce). Total protein was quantified (Bio-Rad) and normalized for protein levels. Proteins were separated by SDS-PAGE electrophoresis using 12% precast Nupage Bis/Tris gels under reducing conditions using MES running buffer (Invitrogen) and transferred onto a nitrocellulose membrane (Amersham Biosciences) using a semidry transfer apparatus (Bio-Rad) and Nupage-MOPS transfer buffer. For immunoblot analysis, membranes were blocked with TBS-Tween containing 5% dry milk and then were incubated with a mouse IgG against human caspase 3 (1:200, Santa Cruz Biotechnology), a mouse IgG against human BAX (1:150, Santa Cruz Biotechnology), a rabbit IgG-HRP against human Actin (1:400, Santa Cruz Biotechnology) and a rabbit IgG against human BCL2 (1:200, Santa Cruz Biotechnology). Bound antibodies were detected by secondary antibodies conjugated to horseradish peroxidase (Santa Cruz Biotechnology). Signal detection was performed by an enhanced chemiluminescence (ECL) detection reagent (Life Technologies). Approximate molecular masses were determined by comparison with the migration of prestained protein standards (Bio-Rad).