Hearing impairment due to Mir183/96/182 mutations suggests both loss and gain of function effects

The microRNA miR-96 is important for hearing, as point mutations in humans and mice result in dominant progressive hearing loss. Mir96 is expressed in sensory cells along with Mir182 and Mir183, but the roles of these closely-linked microRNAs are as yet unknown. Here we analyse mice carrying null alleles of Mir182, and of Mir183 and Mir96 together to investigate their roles in hearing. We found that Mir183/96 heterozygous mice had normal hearing and homozygotes were completely deaf with abnormal hair cell stereocilia bundles and reduced numbers of inner hair cell synapses at four weeks old. Mir182 knockout mice developed normal hearing then exhibited progressive hearing loss. We developed a new bioinformatic tool for creating a causal network connecting transcriptional regulators to their misregulated genes using publicly available data (PoPCoRN) to aid analysis of RNAseq data from the two new mutants. Our transcriptional analyses revealed significant changes in a range of other genes, but surprisingly there were fewer genes with altered expression in the organ of Corti of Mir183/96 null mice compared with our previous findings in Mir96Dmdo mutants which have a point mutation in the miR-96 seed region. This suggests the more severe phenotype of Mir96Dmdo mutants compared with Mir183/96 mutants, including progressive hearing loss in Mir96Dmdo heterozygotes, is likely to be mediated by the gain of novel target genes in addition to the loss of its normal targets.


Background
The microRNAs miR-96, miR-182 and miR-183 are expressed together on a single transcript in sensory cells, including the retina and the hair cells of the inner ear (Weston et al. 2006;Xu et al. 2007). Point mutations in Mir96 cause rapidly progressive hearing loss in the diminuendo mouse mutant (Mir96 Dmdo , (Lewis et al. 2009)) and progressive hearing loss with later onset in human families (Mencia et al. 2009;Solda et al. 2012), and the diminuendo mutation has also been shown to delay maturation of the central auditory system (Schluter et al. 2018). In homozygous Mir96 Dmdo mice, most of the cochlear hair cells die by 28 days after birth. However, this is not the cause of the hearing loss; even before the onset of normal hearing, homozygote hair cells fail to mature both morphologically and physiologically, remaining in their immature state, and heterozygote hair cells show a developmental delay. miR-96 is thus thought to be responsible for coordinating hair cell maturation (Kuhn et al. 2011;Chen et al. 2014).
Overexpression of the three microRNAs also results in cochlear defects and hearing loss (Weston et al. 2018). The complete loss of all mature miRNAs from the inner ear results in very early developmental defects including a severely truncated cochlear duct (Friedman et al. 2009;Soukup et al. 2009). miR-96, miR-182 and miR-183 have also been implicated in other diseases, including glaucoma (Liu et al. 2016), ischemic injury (Cui and Yang 2013; Duan et al. 2019) and spinal cord injury (Ling et al. 2017).
MicroRNAs regulate the expression of many other genes by targeting specific sequences in their mRNAs, leading to transcript destabilisation or translational inhibition. Transcriptome analyses of the Mir96 Dmdo mouse organ of Corti showed that many genes were dysregulated in homozygotes, including several known to be important for hearing which appear to contribute to specific aspects of the diminuendo phenotype (Lewis et al. 2009;Kuhn et al. 2011;Chen et al. 2014;Lewis et al. 2016). However, the diminuendo mutation is a single base pair change in the seed region of the microRNA which is critical for correct targeting, and it is not clear to what extent the diminuendo mutant phenotype is the result of the loss of normal targets of miR-96, and how much is due to the gain of novel targets. We previously suggested that the progressive hearing loss was most likely due to the loss of normal target repression because all three point mutations in mouse and human Mir96 lead to a similar phenotype which seems unlikely if the gain of novel targets is the main mechanism involved.
The regulatory network generated from Mir96 Dmdo expression data (Lewis et al. 2016) includes a number of genes known to be involved in deafness such as Ptprq, Gfi1, Kcna10 and Slc26a5 as well as new candidate genes. Manipulating this network may be a useful therapeutic approach to treating hearing loss due to hair cell dysfunction triggered by a broad range of factors, including both genetic variants and environmental insults. For example, Trp53, Hif1a and Nfe2l2 are in the Mir96 Dmdo network and are involved in cellular responses to stress (Simmons et al. 2009). In order to focus our translational efforts, it is important to understand better the molecular basis of the network. For this reason, we have analysed a second mutation of Mir96 in this report, a double knockout of Mir96 and Mir183, as well as a knockout of the closely linked Mir182 gene, generated through a mouse miRNA knockout program (Prosser et al. 2011). We carried out Auditory Brainstem Response (ABR) and Distortion Product Otoacoustic Emission (DPOAE) tests to characterise their hearing, scanning electron and confocal microscopy to examine their hair cells and synapses, and mRNA sequencing (RNA-seq) to study their transcriptomes. We have also developed a computational tool for connecting regulators with misregulated genes, in order to explore the regulatory networks controlled by miR-96, miR-182 and miR-183. While both new mouse mutants exhibit hearing loss, their electrophysiological and transcriptional phenotypes differ from the diminuendo mouse, with no sign of hearing loss in the heterozygotes, suggesting that the more severe phenotype of Mir96 Dmdo mutants is likely to be mediated by the gain of novel target genes in addition to the loss of its normal targets. We have constructed a new network using only the genes misregulated in both Mir96 Dmdo mutants and the Mir183/96 double knockout.

Mir183/96 and Mir182 knockout mice
Two mouse lines were used in this study; a knockout of Mir182 (Mir182 tm1Hmpr/Wtsi , referred to from here on as Mir182 ko ) and a double knockout of both Mir183 and Mir96 (Mirc40 tm1Hmpr/WtsiOulu , referred to from here on as Mir183/96 dko ), which are only 116bp apart, making it technically challenging to generate two separate knockouts. The mice were generated and maintained on the C57BL/6N genetic background. C57BL/6 mice are known to have age-related hearing loss, partly due to the Cdh23 ahl allele (Noben-Trauth et al. 2003). Higher frequencies are affected first, after four weeks of age, while the lower frequencies remain unaffected up to 6 months old (Li and Borg 1991). We observed a similar pattern in wildtype mice from both the Mir183/96 dko and Mir182 ko lines, which exhibited mild progressive hearing loss at 24-42kHz from 8 weeks old but retained good hearing sensitivity at frequencies between 3-12kHz up to six months old (Fig 1).
For the Mir183/96 dko mice, 43 out of 242 mice (17.8%) from heterozygote by heterozygote matings were homozygous for the Mir183/96 null allele, which is lower than the expected production (25%), suggesting that the absence of Mir96 and/or Mir183 has a small impact on viability (p=0.029, chisquared test). For the Mir182 ko mice, 42 homozygotes out of 152 pups in total (27.6%) were produced from heterozygote by heterozygote matings, which is consistent with the mutation having no impact on viability.

Complete knockout of miRNA expression in the mutant organ of Corti
We carried out qPCR to test the expression levels of the three microRNAs in the organs of Corti of wildtypes, heterozygotes and homozygotes of each knockout. In Mir183/96 dko homozygotes, there was no detectable expression of Mir183 or Mir96. Likewise, in Mir182 ko homozygotes there was no detectable expression of Mir182 (Appendix Figure S1). The levels of expression in heterozygotes of both knockouts was variable, as was the expression of Mir182 in Mir183/96 dko homozygotes, and Mir183 and Mir96 in Mir182 ko homozygotes. It is likely that this is because we lack a proper microRNA control for sensory tissue in the inner ear. We used Mir99a, which is expressed in almost all cell types in the cochlea, including hair and supporting cells (Friedman et al. 2009), but because Mir183, Mir182 and Mir96 are also expressed in hair cells, it is possible that the mutant alleles affect the expression of Mir99a in hair cells, making it an unreliable calibrator between wildtype, heterozygote and homozygote. A better calibrator would be a microRNA expressed only in supporting cells and not in hair cells.

Impaired auditory responses in homozygotes but normal thresholds in heterozygotes
Mir183/96 dko homozygous mice were profoundly deaf, with most showing no response at the highest sound level tested (95dB SPL) at any of the ages tested (14 days to six months old). In contrast, Mir182 ko homozygotes exhibited only a mild hearing loss starting at higher frequencies and progressing with age to lower frequencies (Fig 1). The ABR thresholds of heterozygotes were normal at all ages tested (Fig 1, see Appendix Figures S2, S3 for individually plotted traces). ABR waveforms of Mir183/96 dko heterozygotes and Mir182 ko homozygotes were also similar to those of wildtype littermates at the equivalent sound pressure level above threshold (sensation level, SL) (Appendix Figure S4). We measured DPOAEs at 8 weeks old and found no difference in the amplitudes or thresholds between wildtype and heterozygous Mir183/96 dko mice, while homozygotes had severely abnormal responses (Appendix Figure S6). Mir182 ko homozygotes had raised DPOAE thresholds at high frequencies compared to wildtypes (Appendix Figure S6), which matched the difference in their ABR thresholds at 8 weeks (Fig 1). Mir182 ko mutant mice showed no sign of a vestibular defect (circling, head-bobbing or hyperactivity) up to six months old. However, Mir183/96 dko homozygotes did show increasing incidence of hyperactivity with age (Appendix Figure S5).
Heterozygous Mir183/96 dko mice recover normally from noise exposure As heterozygous Mir183/96 dko mice showed no auditory deficit, in contrast to the hearing loss seen in diminuendo heterozygotes and in humans carrying one mutant MIR96 allele, we asked if these heterozygous mice might be more sensitive to noise-induced damage. One day after noise exposure (2 hours of 8-16kHz octave-band noise at 96dB SPL) at 8 weeks old, both Mir183/96 dko heterozygous and wildtype mice showed a marked increase in thresholds at 12kHz and above compared to unexposed control littermates (Fig 2). Three days after noise exposure the 12kHz thresholds had recovered, but there was still a noticeable elevation at higher frequencies. By 7 days after exposure, all thresholds had recovered completely (Fig 2). We measured the amplitude of wave 1 of the ABR waveform to look for a reduced neural response, which has been reported in CBA/CaJ mice after noise exposure (Kujawa and Liberman 2009) and is thought to be due to neuronal loss in the cochlea, but no difference was observed (Fig 2). At 24kHz, we observed a much greater effect, such that in most animals wave 1 was too poorly defined to measure the amplitude one day after exposure. However, both wildtype and heterozygote wave 1 amplitudes had recovered by 28 days after exposure, and did not look any different to mice which had not been exposed (Fig 2; Appendix Figure S7).
Since there was a significant difference in the higher frequencies between the unexposed heterozygotes and wildtypes at 8 weeks old (Fig 2), but we did not see any difference in our original ABR tests (Fig 1), we compared the ABR thresholds from all mice tested at 8 weeks old (Appendix Figure S8). At high frequencies (30-42kHz) the thresholds were very variable in both heterozygotes and wildtypes. The differences in the means at 36kHz and 42kHz are statistically significant (P = 0.001, P=0.000; mixed linear model pairwise comparison), but given the variability we suggest that this is not biologically relevant.

Hair cells and innervation
We examined the organ of Corti using scanning electron microscopy and found that hair cells in homozygous Mir183/96 dko mice were severely affected at four weeks old (Fig 3), with many hair bundles missing entirely. Where present, the stereocilia bundles of both outer and inner hair cells show splaying and fusion. The inner hair cells of Mir183/96 dko heterozygotes are unaffected, but the outer hair cells' upper surface appear slightly smaller and rounder in shape than the wildtype outer hair cells (Fig 3). Their stereocilia bundles also appear smaller and more rounded than normal, with more pronounced tapering in height and overlap of shorter stereocilia rows with taller rows towards the two ends of each bundle.
Four week old Mir182 ko heterozygotes and homozygotes showed no abnormalities of hair cells by scanning electron microscopy at either the 12kHz or 24kHz regions (Appendix Figure S9), corresponding to their normal ABR thresholds at that age. At eight weeks old, when hearing loss is evident at 24kHz and higher (Fig 1), we also saw no systematic differences between wildtypes, heterozygotes and homozygotes (Fig 4).
The distribution of unmyelinated neurons appeared normal in both mutants using antineurofilament labelling (Appendix Figure S10). Synapses were examined using anti-Ribeye antibody to mark presynaptic ribbons and anti-GluR2 to mark postsynaptic densities, and a significant reduction was found in the number of colocalised pre-and postsynaptic markers in Mir183/96 dko homozygotes (Fig 5, p=0.036, Wilcoxon rank sum test). No difference in synapse counts was observed in Mir182 ko homozygotes.

Transcriptome analysis reveals misregulation of gene expression in mutants
To investigate the impact of the Mir182 ko and Mir183/96 dko mutations on gene expression we carried out RNAseq of isolated organ of Corti preparations from postnatal day (P)4 homozygotes and sexmatched littermate wildtype controls. This age was chosen to ensure all hair cells were still present and to facilitate comparison with our previous transcriptome data from Mir96 Dmdo mice (Lewis et al. 2009). Thirty-four genes were identified as significantly misregulated (FDR < 0.05) in Mir183/96 dko homozygotes, 22 upregulated and 12 downregulated compared with wildtype littermates. Many of the upregulated genes have sequences complementary to either the miR-96 seed region or the miR-183 seed region in their 3'UTRs (Table 1). Of this list of 34 genes, only Hspa2, Ocm, Myo3a, Slc26a5, Slc52a3, St8sia3 and Sema3e were previously found to be misregulated in Mir96 Dmdo mice at P4 and/or P0 (Lewis et al. 2009;Lewis et al. 2016), and in each case the misregulation is in the same direction (Table 1). We tested 19 genes of the 34, selected because they were reported to show a large difference in expression levels between sensory and non-sensory cells in the organ of Corti (http://www.umgear.org; Cai et al. 2015;Elkon et al. 2015). All but three were confirmed correct (Table 1, Appendix Figure S11); those that were not confirmed by qPCR were either not misregulated or failed the significance test, and no genes were significantly misregulated in the opposite direction.
It is possible that the difference between the transcriptomes of Mir183/96 dko knockout mice and Mir96 Dmdo mice is due to the different backgrounds. We looked for differences in predicted targets in the 3'UTRs of the C57BL/6NJ and C3H/HeJ genome sequences (Adams et al. 2015), which were the closest genome sequences to our mutant backgrounds available, and found that 1585 genes had the same number of seed matches, and 13 genes had seed matches in both strains, but not the same number of matches. 36 genes had no seed matches in the C57BL/6NJ sequence but had one or more match in the C3H/HeJ sequence, and 47 genes had no seed matches in the C3H/HeJ sequence, but one or more in the C57BL/6NJ sequence (Appendix Table S1). None of these 83 genes were in the list of miR-96 targets used for network creation.
Three genes were found to be significantly upregulated (FDR < 0.05) in the Mir182 ko homozygotes by RNA-seq, one of which, Ppm1l, has sequences complementary to the seed region of miR-182 (Table   1). The other two, Grp and Ccer2, were also upregulated in the Mir183/96 dko homozygotes. No genes were significantly downregulated in Mir182 ko . We tested the upregulated genes by qPCR, and also tested Slc26a5 and Ocm (which were strongly downregulated in Mir96 Dmdo homozygotes (Lewis et al. 2009) and found that only the upregulation of Grp was confirmed. Ccer2 and Ppm1l were upregulated but not significantly, and Slc26a5 and Ocm were downregulated but again, not significantly (Table 1, Appendix Figure S11).  figure 6A), and the sequence complementary to the seed region of miR-183 has a small peak towards the centre of the graph (dark blue line, figure 6A).
There were no miRNA seed region heptamers enriched in the Mir182 ko gene list ( figure 6B), but the TATTTAT heptamer which is enriched in the Mir182 ko downregulated genes (yellow line, figure 6B) resembles a portion of an AU-rich element. These are 50-150bp sequence elements found in 3'UTRs, and are typically involved in mRNA destabilising via a deadenylation-dependent mechanism (reviewed in (Chen and Shyu 1995;Barreau et al. 2005)). AU-rich elements, and TATTTAT in particular, are enriched in 3'UTR regions which have multiple RNA-binding protein binding sites, and it has been suggested that RNA-binding proteins compete with the RISC complex to bind to miRNA target sites within these regions (Plass et al. 2017). It is possible that this TATTTAT signal, which is enriched in the genes downregulated in Mir182 ko homozygote hair cells, is the result of a change in the binding of RNA-binding proteins in the absence of miR-182.

No evidence for differential splicing
Splicing patterns can be affected by a microRNA if it targets a splicing factor, as has been shown for miR-133a and nPTB (Boutz et al. 2007), so we analysed the Mir183/96 dko and Mir182 ko RNA-seq data for evidence of alternative splicing events. Tools which test RNA-seq data for differential splicing of mRNAs are not as well established as those for measuring differential expression, and there is great variability in the results depending on which method is used (Liu et al. 2014). We therefore used three different tools to test for differential splicing, which all take a different approach. Cuffdiff (from the Cufflinks package (Trapnell et al. 2010)) maps reads to individual transcripts in order to estimate the relative abundance of each isoform. This can introduce uncertainty in the case of reads which map to exonic regions shared by more than one transcript. JunctionSeq (Hartley and Mullikin 2016), is a count-based tool which tests read counts belonging to exonic regions specific to individual transcripts, avoiding the problem of shared regions but potentially missing isoforms which have no unique exons. Leafcutter, a very recently developed algorithm, tests for differential splicing by using reads which span exon-exon junctions, in effect focussing on the excised intron (Li et al. 2016).
All three tools detected significant differential splicing in Mir183/96 dko homozygotes compared to their wildtype littermates (Appendix Table S2), but only three genes were identified by more than one tool; Ppp3cb and Zfp618 (by Leafcutter and JunctionSeq), and Il17re (by Leafcutter and Cuffdiff).
One of the advantages of Leafcutter and JunctionSeq is that they identify the sites at which they detect differential splicing occurring, which allowed us to test their results. We designed primers to cover the junctions and/or exons which were highly differentially spliced (15 genes; appendix Table   S3) and sequenced PCR reactions from cDNA made from organ of Corti RNA from Mir183/96 dko homozygotes and wildtypes. We found no evidence of the differential splicing detected by either tool, but evidence of a novel isoform of Stard9 was detected by JunctionSeq. The novel isoform includes an exon (ENSMUSE00001437951) previously assigned to a transcript subject to nonsensemediated decay (transcript ENSMUST00000140843), but the exons following ENSMUSE00001437951 in our results belong to the protein-coding transcript (ENSMUST00000180041). Exon ENSMUSE00001437951 is 72bp long, so does not introduce a frameshift, and its inclusion may result in a functional Stard9 protein with 24 extra amino acids. The novel splicing was present in both wildtype and homozygous Mir183/96 dko RNA (Appendix Figure S12).
JunctionSeq did not predict any significant differential splicing in Mir182 ko homozygotes. Leafcutter and Cuffdiff identified some genes with differential splicing (Appendix Table S2), but there were no genes common to both predictors. Two genes, Ube4a and Nrxn1, were predicted to have alternative splicing present in the homozygote and not in the wildtype, but when we sequenced their mRNA from Mir182 ko homozygotes and wildtypes, we found no differences in the splicing pattern.

Immunohistochemistry confirms downregulation of Ocm
We carried out antibody stains on sections from the inner ear at P4, to check for the presence of Ocm (oncomodulin) and Slc26a5 (prestin) protein in the hair cells. Ocm staining is faint at P4, stronger in the basal turn of the cochlea, but prestin staining is clearly visible at that stage in wildtypes. We observed prestin staining in Mir183/96 dko homozygotes, but no stain for Ocm, while both proteins were present in the wildtype littermate controls (Appendix Figure S13). Although immunohistochemistry is not a quantitative technique, this correlates with the qPCR results, which showed that Ocm RNA was nearly absent in Mir183/96 dko homozygotes, while Slc26a5 RNA levels were about 30% of wildtype levels (Appendix Figure S11)

. Ocm and Prestin staining was visible in
Mir182 ko homozygotes and their wildtype littermate controls (Appendix Figure S13).

Network analysis suggests upstream regulators
We took three approaches to network analysis: Ingenuity Pathway Analysis, WGCNA and a new tool, PoPCoRN. First, we used Ingenuity Pathway Analysis to construct putative causal networks using the misregulated genes in each dataset with FDR<0.05, as follows. Upstream regulators which could explain the observed misregulation were identified through the Ingenuity Knowledge Base, a manually curated set of observations from the literature and from third-party databases. The activation or inhibition of each upstream regulator was calculated based on the observed misregulation of its target genes. Causal networks were then constructed to link the upstream regulators through a "root" regulator further upstream, again using the Ingenuity Knowledge Base (described in detail in (Kramer et al. 2014)). The resulting networks were filtered based on their activation z-score, which is a measure of the match of observed and predicted misregulation patterns. The z-score is not just a measure of significance but also a prediction of the activation of the upstream root regulator (Kramer et al. 2014); scores below -2 are classed as significant by Ingenuity Pathway Analysis and indicate predicted downregulation or reduced activation of the root regulator, while scores above 2 are also considered significant and indicate predicted upregulation or increased activation of the root regulator. For the Mir183/96 dko homozygotes, only four upstream regulators were given a z-score < -2, three more had z-scores < -1.7 and one had a z-score > 1.7.
These eight upstream regulators, Bdnf, Cshl1, H2afx, Kcna3, Kit, Myocd, Prkaca and Pth, explain between them the misregulation of 14 genes (Fig 7), although they themselves are not misregulated (from the RNAseq data). Most of the genes are predicted to be inhibited, so they cannot be directly linked to miR-96 (although Bdnf is a known target of miR-96 (Li et al. 2015), which does not fit the predicted inhibition). Only Kcna3 is predicted to be activated (orange in Fig 7), and it does not have any matches to the miR-96 or miR-183 seed regions in its 3' UTR. Of these eight potential regulators, Bdnf and Kit are both known deafness genes (Deol 1970;Agerman et al. 2003). There were too few genes significantly misregulated in the Mir182 ko RNA-seq data to obtain any causal networks with a significant activation z-score, but Ret and Adcyap1 were identified as immediate upstream regulators of Grp.
Gene clustering using weighted gene correlation network analysis (WGCNA) suggests further regulators Weighted gene correlation network analysis is a method of analysing transcriptome data to cluster genes into modules based on their expression across a number of individual samples without reference to the sample traits (Zhang and Horvath 2005). We used a Pearson correlation to cluster genes across all 24 samples (Mir183/96 dko and Mir182 ko ) and obtained 29 consensus modules (including the reserved "grey" module, which consists of genes outside of all the other modules) (Appendix Table S4).
These consensus modules are groups of genes with highly correlated expression profiles. For further analysis, we obtained each module's eigengene, which represents all the genes within the module, and we calculated the correlation of each eigengene with the traits of the mice used for RNAseq (Appendix Figure S14). We then clustered the eigengenes to identify meta-modules where eigengenes were highly correlated with each other (Appendix Figure S15). Of the consensus module eigengenes, three were highly correlated and clustered with the wildtype vs. Mir183/96 dko homozygote trait (green, black, royal blue, Appendix Figure S15A). This means that the expression levels of the genes of the green, black and royal blue modules are correlated with each other and the genotypes of the Mir183/96 dko mice. One module was highly correlated and clustered with the wildtype vs. Mir182 ko homozygote trait (salmon, Appendix Figure S15B) We chose nine modules for further exploration which had significant correlation with the wildtype vs. Mir183/96 dko homozygote or wildtype vs. Mir182 ko homozygote traits (Appendix Figure S14; black, green, royal blue, salmon, purple, grey60, light green and dark green modules). We also investigated the blue module, which contained the most differentially expressed genes (13, Appendix Table S4)). We carried out enrichment analysis using PANTHER v14 ( We also investigated potential transcription factors which could be responsible for the altered regulation of the genes within each module using oPOSSUM (Kwon et al. 2012), which detects overrepresented transcription factor binding sites in a set of genes. We obtained 29 trancsription factor binding site profiles across the 9 modules with Z and Fisher scores above our chosen threshold of the mean + 1 standard deviation (Appendix Figure S16, Appendix Table S4) (Kwon et al. 2012). Several of the profiles were shared between modules (Appendix Figure S17). Transcription factors implicated in the black, green, blue and light green modules include one deafness gene (Foxi1 (Hulander et al. 1998)) and several genes with miR-183/96/182 seed region matches, although none have yet been experimentally shown to be targets of any of the three microRNAs (Appendix Table S7).
Transcription factors implicated in the purple, royal blue and dark green modules also include genes with seed region matches, and Myc, which is also a deafness gene (Wei et al. 2007), is a potential regulator of the genes in the grey 60 module (Appendix Table S7). None of the transcription factors were found to be misregulated in the RNAseq data.

The PoPCoRN tool extends the networks
The IPA networks did not connect the misregulated genes to any of the three microRNAs, probably because each microRNA controls a regulatory cascade, and the IPA causal networks have a maximum depth of 3, which may not be enough to reach the level of direct targets. The WGCNA analysis identified modules of co-expressed genes and highlighted transcription factors which may be involved, but did not connect these directly to either the microRNAs or the misregulated genes.
We therefore developed the PoPCoRN (prediction of potential causal regulatory networks) tool to make use of publicly available regulatory data to connect misregulated genes to a known top-level regulator, and used this to analyse the Mir183/96 dko and Mir182 ko datasets. PoPCoRN requires a list of misregulated genes and a list of genes at which to stop, the upstream regulators beyond which no connections will be explored. We used the significantly misregulated genes (FDR < 0.05) from each RNA-seq experiment (Table 1) and used the relevant microRNAs as the genes at which to stop. The PoPCoRN tool is freely available from github (https://github.com/moraglewis/PoPCoRN).
The resulting Mir183/96 dko network consists of 114 genes and 416 links. Thirty misregulated genes have been included, 13 of them predicted targets of either miR-183 or miR-96 or both. Two misregulated genes were left out completely due to having no known regulatory links in our compiled potential links (Ccer2, and Rn7s1) (Fig 8A). We carried out a network analysis using the Cytoscape Network Analyser tool to calculate the degree and betweenness centrality of each node (gene). A node's degree is the number of edges connecting it to other nodes, and its betweenness centrality measures how important it is for connecting distant parts of the network. The node with the highest degree is Trp53, with 23 edges. Most nodes have 9 or fewer edges ( Fig 8A), but 34 have 10 or more edges, including seven which are direct targets of miR-96, miR-183 or both (listed in Table 2). None of these 34 nodes with 10 or more edges are known to be misregulated in Mir183/96 dko homozygotes, but eight are known deafness genes (Fos (Paylor et al. 1994), Foxo3 Rest (Nakano et al. 2018) and Tnf (Oishi et al. 2013)). Among these 34 nodes are Trp53, Sp1, Tgfb1, Tnf and Fos, which were also genes of interest in the manually-created network based on the Mir96 Dmdo microarray data (Lewis et al. 2016).
We also carried out the same analysis using the microarray and qRTPCR data from Mir96 Dmdo (Lewis et al. 2016) (159 input genes with adjusted P < 0.05). The resulting network has 464 genes and 2561 links, and includes 127 of the 133 misregulated genes (Fig 8B), in contrast to the network created manually in our previous work, which consisted of 103 genes with 343 links, and included only 57 of the misregulated genes (Lewis et al. 2016). In this new Mir96 Dmdo network, the node with the highest degree is E2f2, with 76 edges. There are 28 nodes with 31 or more edges, all of which are also part of the Mir183/96 dko network. Alk, Foxo1, Foxo3, Rad51, Bdnf, Irs1 and Nr3c1 ( Fig 8C). The three nodes with the highest edge count are Rest (12), Trp53 (11), and Fos (10). Of these, Rest is a known deafness gene (Nakano et al. 2018) and Fos was originally predicted to be downregulated in the manually created Mir96 Dmdo network, and this was confirmed by qPCR carried out on Mir96 Dmdo homozygotes (Lewis et al. 2016).
The Mir182 ko network is very small, being based on three input genes: Ppm1l, Ccer2 and Grp. Ccer2 could not be included, due to the lack of any known regulators, and Ppm1l is a predicted target of miR-182. Multiple potential pathways link Grp to miR-182, but all of them work through Hnf4a, which upregulates Grp (Fig 8D).
To validate the network analyses, we carried out qRTPCR on 11 genes from the Mir183/96 dko network and two from the Mir182 ko network (Fig 8). Our choice was based on several factors, including their importance within the network, their edge count and close links to important genes (eg Ocm and Slc26a5). Fos and Trp53 were also predicted to be involved by the IPA analysis (Fig 7).
However, we found all the genes had very variable expression levels in homozygotes when compared to littermate wildtypes (Appendix Figure S11). It's possible that this is the result of expression in different areas of the inner ear. We controlled for the quantity of sensory tissue using Jag1, but the inner ear is a complex organ with many cellular subtypes, and if a gene is expressed in multiple cell types, controlling for overall tissue quantity (using Hprt) and sensory tissue quantity (using Jag1) may not be sufficient to permit detection of small changes in mRNA levels in the sensory tissue. In addition, qRTPCR can only detect mRNA levels, and can't directly measure protein levels or protein activity.

Discussion
Mir183/96 dko mice have a less severe phenotype than Mir96 -Dmdo mice Mice heterozygous for the Mir96 Dmdo point mutation exhibit early-onset rapidly progressive hearing loss; even at P15 they have very raised thresholds (Kuhn et al. 2011). We initially concluded that this was more likely to be due to haploinsufficiency than to the gain of novel targets, because humans heterozygous for different point mutations also had progressive hearing loss (Mencia et al. 2009 (Lumayag et al. 2013;Fan et al. 2017).
Stereocilia bundles and innervation in Mir183/96 dko mice and Mir96 Dmdo mice The hair cells of Mir96 Dmdo homozygous mice are present at four days old but appear abnormal, and by 28 days old have degenerated almost completely (Lewis et al. 2009). In Mir183/96 dko homozygotes, however, some hair cell stereocilia bundles are still visible at P28, although they are severely disorganised (Fig 3). In Mir183/96 dko heterozygotes, IHC stereocilia are mostly normal and the OHC stereocilia bundles appear to be slightly rounded in arrangement (Fig 3), but this is not reflected in their ABR thresholds or DPOAE responses, which are normal (Fig 1 and Fig S5). Mir96 Dmdo heterozygotes, which are deaf by P28, have a more severe phenotype, with fused stereocilia and hair cell degeneration as well as misshapen stereocilia bundles in outer hair cells, and smaller stereocilia bundles in inner hair cells (Lewis et al. 2009).
Heterozygote Mir183/96 dko mice showed no differences in innervation or synapse counts (Appendix Figure S10, Fig 5). Similar to the physiological phenotype, the structural phenotype of Mir183/96 dko heterozygotes is much less severe than that of Mir96 Dmdo heterozygotes. Sema3e (Table 1). Some of the differences will be due to certain genes not being represented on the microarray assay when it was carried out, such as Ccer2, and some may be the result of the different genetic background. We aimed to minimise this by use of the sex-matched wildtype littermates as controls, but it isn't possible to entirely eliminate the effect of genetic background; there are 83 genes which lack miR-96 seed region matches in one background compared to the other (Appendix Table S1), and any of those may be contributing to the phenotype. Much of the difference, however, is likely to be due to the loss of miR-183 in the Mir183/96 dko homozygotes and the presence of the mutant miR-96 in Mir96 Dmdo homozygotes. We were unable to confirm any of the predicted differential mRNA splicing in Mir183/96 dko homozygotes, suggesting that differential splicing is unlikely to play a large part in the mutant phenotypes.

Network analyses of the Mir183/96 dko transcriptome suggest multiple potential regulators
We took three approaches to find intermediate regulators, and obtained 8 upstream regulators from the Ingenuity Pathway Analysis causal network analysis, 29 transcription factor profiles from the WGCNA module analyses (Appendix Table S7), and 34 highly connected nodes from the PoPCoRN network construction (Table 2). There are no genes suggested by all three approaches, but Cebpa is a highly connected PoPCoRN node and one of the transcription factors which is implicated in the WGCNA module analysis, and Kit, which is a deafness gene (Deol 1970), is predicted to be an upstream regulator by Ingenuity, as well as being a highly connected node in the PoPCoRN network.
Other genes of particular interest in our network analyses are Zeb1, Rest, and Ikzf2, being all three deafness genes with a particular role to play in specification and maturation of hair cells. Zeb1 mRNA is depleted in the sensory epithelial cells of the inner ear by miR-200b, and a mutation causing upregulation of Zeb1 in the inner ear results in auditory and phenotypic defects (Hertzano et al. 2011). Zeb1 has also been shown to be a direct target of miR-96 in a different model of the epithelial-mesenchymal transition (Li et al. 2014). Rest is a neuronal gene repressor and is downregulated in hair cells by both transcriptional and splicing mechanisms; its upregulation results in hair cell degeneration and auditory and vestibular defects in mice (Nakano et al. 2018). It is predicted to be upregulated/activated by the PoPCoRN network (figure 8A,D). Ikzf2 (helios) is required for outer hair cell maturation, and is known to regulate both Ocm and Slc26a5 (Chessum et al. 2018). It is predicted to be downregulated in the Mir183/96 dko network ( Figure 8A). PoPCoRN network and were also listed by the WGCNA transcription factor analysis. The network created using the seven genes misregulated in both the Mir183/96 dko and Mir96 Dmdo transcriptome datasets implicates several potential targets for miR-96 in the inner ear (Fig 8), including three known deafness genes: Foxo3, which has been linked to auditory neuropathy in mice (Gilels et al.

2015).
The gain of novel targets plays an important role in the phenotype resulting from a point mutation in Mir96.
We originally hypothesised that the major effect of the Mir96 Dmdo mutation on hearing was most likely mediated by a reduced level of downregulation of the normal target mRNAs (Lewis et al.

A model for mechanisms of action of mutant microRNAs
It has been suggested that miRNAs act in two ways; first, they repress targets to prevent translation, resulting in mutually exclusive expression of the miRNA and the target, and second, they buffer transcriptional noise, in which case the miRNA and its targets are co-expressed in the same cell (Hornstein and Shomron 2006). Our approach should be able to detect both effects, but if a target is highly expressed in the non-sensory epithelial cells, the difference in expression between wildtype and homozygote due to derepression in the hair cells may not be detectable. Our current transcriptome data, therefore, is more likely to highlight targets which are being buffered rather than targets which are completely repressed by miR-96 (Hertzano et al. 2011). This may explain why likely target genes like Zeb1 and Nr3c1 have not been found to be significantly misregulated in our transcriptome data. It may also explain the variability of the network genes we tested (Fig S11E,F (Fig 9).
We suggest that the consistent misregulation of Ocm, Slc26a5, Myo3a, Sema3e and Slc52a3 observed in both the Mir96 Dmdo and Mir183/96 dko homozygotes is the result of the lack of repression of genes that would usually not be expressed in hair cells at all, for example Zeb1, Foxo1, Nr3c1 and Irs1. The links between these targets and the downstream genes are yet to be discovered, but we suggest Ikzf2 and Fos are likely to be involved (Fig 9A, B). There is one gene which is upregulated in the Mir183/96 dko homozygote, has two matches to the miR-96 seed region in its 3'UTR and is strongly expressed in the cochlear duct but not in the hair cells, and that is Eln There are multiple genes which are known or potential miR-96 targets, upregulated in miR-96 mutants, and expressed in hair cells, such as Hspa2, St8sia3, Fadd and Grk1. We suggest these are examples of genes whose expression is buffered by miR-96, not fully repressed but maintained at a consistent level. The loss of this buffering results in variable expression and transcriptional noise, not just of these genes (which will all be upregulated to varying degrees) but also of any genes which they regulate. This transcriptional noise is less likely to be consistent (at the gene level) between mice but will nonetheless contribute to the degraded functionality of the hair cell (Fig 9A, B).
Finally, genes which are novel targets of a mutant microRNA have the potential to fulfil both roles. In the case of Mir96 Dmdo , Ptprq is downregulated and it's possible this is due to the mutant microRNA, since Ptprq bears a complementary match to the Mir96 Dmdo seed region in its 3' UTR. A reduction in Ptprq in the hair cells will contribute directly to the failure of the hair cells to mature properly. Other novel targets may also be directly affecting the hair cells, or may be adding to the overall transcriptional noise, or both ( Fig 9C).

The role of Mir183 in hearing remains unclear
The lack of a Mir183-specific mutation means that the effect of miR-183 alone is difficult to ascertain from our data. The relatively low mid-range peak for the miR-183 seed region in our Sylamer analysis of the RNA-seq data (Fig 6) implies that the lack of miR-183 has less of a global effect on the transcriptome than the lack of miR-96, and the PoPCoRN network analysis reflects this; apart from its predicted targets, there are no downstream genes whose misregulation can be attributed to miR-183 alone (Fig 8A). In the network generated from the Mir183/96 data, the only miR-183 targets which regulate downstream genes are also miR-96 targets.
Mice lacking Mir182 display mild hearing loss with no obvious changes in stereocilia bundles or hair cell innervation The Mir182 ko heterozygotes have audiograms which resemble the wildtype, while the homozygotes exhibit mild hearing loss at the higher frequencies (Fig 1). No difference in hair cell stereocilia bundles, innervation or synapse counts was observed between the Mir182 ko wildtype, heterozygous and homozygous mice at P28 (Appendix Figures S9, S10, Fig 5), but at that age Mir182 ko homozygotes still have normal hearing. However, even at P56, when Mir182 ko homozygotes exhibit high frequency hearing loss, the hair cells appear unaffected (Fig 4). This is a completely different phenotype from either the Mir96 Dmdo or the Mir183/96 dko mice and may explain the paucity of significantly misregulated genes in the RNA-seq data from the P4 organ of Corti. This may be too early to see much effect of the absence of miR-182 on the transcriptome. The lack of significantly enriched heptamers corresponding to the miR-182 seed region in the Sylamer analysis for miR-182 also supports this hypothesis. The network drawn from the Mir182 ko transcriptome may show the start of the misregulation cascade due to the microRNA knockout, and the transcription factors suggested by the WGCNA and IPA analyses may be involved, but further studies at a later age would be required for confirmation. None of the predicted differential mRNA splicing in Mir182 ko homozygotes was confirmed.

Relevance to human MIR96 mutations
In this study we have shown that the phenotype of mice lacking miR-96 entirely is less severe than that of mice carrying a point mutation in the miR-96 seed region. Mice heterozygous for the Mir183/96 dko allele display no auditory phenotype when compared to wildtypes, while Mir96 Dmdo heterozygotes have rapid, severe progressive hearing loss. In our first study of the diminuendo point mutation, we concluded that the phenotype caused by the mutant miR-96 was most likely the result of loss of normal miR-96 targets, because all three mutations described (two in humans and one in the mouse) resulted in hearing loss (Lewis et al. 2009;Mencia et al. 2009). Our data from the current study, however, suggest that the gain of novel targets also plays an important role in the phenotype caused by mutations in miR-96. This has important implications for understanding the effect of mutant microRNAs in the human population. So far all three reported human mutations in miR-96 have been point mutations, two in the seed region and one in the stem region of the pre-miRNA (Mencia et al. 2009;Solda et al. 2012), and while all the individuals carrying those point mutations have exhibited some degree of progressive hearing loss, the phenotypes differ between people, as do the phenotypes of the mice carrying the Mir96 Dmdo mutation and the Mir183/96 dko allele.
Determining the misregulated pathways common to all Mir96 mutations will be important not only for furthering our understanding of the role of this microRNA in hair cell development but also for developing therapeutic interventions, not only for people with MIR96 mutations but also more generally for hearing loss associated with hair cell defects.

A new bioinformatic analysis tool
Our network tool (PoPCoRN) makes use of publicly available data to draw a causal network of potential regulatory interactions which could explain the observed misregulation. Similar approaches have been described previously (Pollard et al. 2005;Chindelevitch et al. 2012;Kramer et al. 2014;Fakhry et al. 2016), but they address the case where the upstream regulator or regulators remain unknown, for example when comparing samples from healthy and diseased tissues, and one of their main aims is to identify potential upstream regulators. Our implementation asks a different question, exploring the downstream regulatory cascade of a known regulator. In the case of miR-96, and potentially many other regulators involved in disease, the genes involved in mediating their effect are candidate therapeutic targets. and wildtype littermates was tested using the Auditory Brainstem Response, following the protocol described in . Animals were sedated using a ketamine/xylazine mix (10mg ketamine and 0.1mg xylazine in 0.1ml per 10g body weight) and recovered where possible using atipamezole (0.01mg atipamezole in 0.1ml per 10g body weight) to enable recurrent testing of the same cohort of mice. All injections were intraperitoneal. Responses were recorded from three subcutaneous needle electrodes placed one over the left bulla (reference), one over the right bulla (ground) and one on the vertex (active). We used a broadband click stimulus and 3kHz, 6kHz, 12kHz, 18kHz, 24kHz, 30kHz, 36kHz and 42kHz pure tone frequencies, at sound levels from 0-95dB, in 5dB steps. 256 stimulus presentations per frequency were carried out per frequency and sound level, and responses were averaged to produce the ABR waveform. The threshold for each stimulus, the lowest intensity at which a waveform could be distinguished visually, was identified using a stack of response waveforms. Mice were tested at 14 days old (P14), P21, P28 ±1 day, P56 ±2 days, P90 ±2 days and P180 ±3 days. Any hyperactivity was noted prior to anaesthesia. Wave 1 amplitudes were calculated using ABR Notebook software (courtesy of MC Liberman, Harvard Medical School/Massachusetts Eye and Ear).

Methods
Distortion Product Otoacoustic Emission (DPOAE) measurements. We measured DPOAEs in mice aged 8 weeks old, anaesthetised with an intraperitoneal injection of 0.1ml / 10g of a solution of 20% urethane. Experiments were performed using Tucker Davis Technologies (TDT) BioSigRZ software driving a TDT RZ6 auditory processor and a pair of TDT MF1 magnetic loudspeakers. Signals were recorded via an Etymotic ER-10B+ low-noise DPOAE microphone. Stimulus tones (f1 & f2) were presented and microphone signals recorded via a closed-field acoustic system sealed into the auditory meatus of the mouse. Stimulus tones were presented at an f2:f1 ratio of 1.2. f2 tones were presented at frequencies to match ABR measurements (6,12,18,24,30 and 36 kHz). f1 was presented at levels from 0-85 dB in 5dB steps. f2 was presented at 10 dB below the level of f1. The magnitude of the 2f1-f2 DPOAE component was extracted from a fast Fourier transform of the recorded microphone signal and plotted as a function of f2 level. For each f2 level, the 20 spectral line magnitudes surrounding the 2f1-f2 frequency were averaged to form a mean noise floor estimate for each measurement. DPOAE threshold was defined as the lowest f2 stimulus level where the emission magnitude exceeded 2 standard deviations above the mean noise floor.
Noise exposure. Wildtype and heterozygous Mir183/96 dko mice (P55±1 day) were subjected to an 8-16kHz octave-band noise at 96dB SPL for two hours while awake and unrestrained in separate small cages within an exposure chamber designed to provide a uniform sound field (for chamber details, see (Holme and Steel 2004)). Band pass noise was generated digitally using TDT RPvdsEx software, converted to an analogue signal using a TDT RZ6 auditory processor, and amplified using a Brüel and Kjaer Type 2716C power amplifier. It was delivered to a compression driver (JBL 2446H, Northridge, CA) connected to a flat front biradial horn (JBL 2380A, Northridge, CA) secured to the roof of the sound box. ABRs were carried out the day before, and 1, 3, 7, 14 and 28 days after the noise exposure. Unexposed littermates were used as controls and went through the same set of ABR measurements.
Genotyping. Mir183/96 dko knockout mice were genotyped by PCR analysis using primers spanning the introduced deletion (Appendix Table S3). The wildtype band is 841bp and the mutant band 645bp. Mir182 ko mice were genotyped in a similar fashion with one of two primer sets (Appendix Table S3). The wildtype band for the first set is 495bp and the mutant band 457bp. The wildtype band for the second set is 247bp, and the mutant band is 209bp.
Scanning electron microscopy. The inner ears of wildtype, heterozygote and homozygote mice at P28 (Mir183/96 dko , Mir182 ko ) and P56 (Mir182 ko ) were fixed in 2.5% glutaraldehyde in 0.1M sodium cacodylate buffer with 3mM CaCl 2 at room temperature for two hours. Cochleae were finely dissected in PBS and processed according to the OTOTO method (Hunter-Duvar 1978). Samples were dehydrated using an ethanol series, critical point dried and mounted for examination. Low resolution images were taken to identify the 12kHz region of the cochlea (using the frequency-place map described by Müller et al (Muller et al. 2005)). Higher-resolution images were taken using a JEOL JSM 7800 Prime scanning electron microscope under a standard magnification (60x) to show the whole organ of Corti, and at higher magnifications to examine hair cell rows (2000x)  region, then a close-up image (2.5x optical zoom) was taken from both extremes of the low magnification image to ensure no synapses were counted twice. Synapses were counted from the maximum projection of the close up images using the FIJI plugin of ImageJ, and divided by the number of hair cell nuclei visible in the field of view (between 5 and 13, depending on which microscope was used) to obtain the number of synapses/hair cell. Where synapses were counted in both ears from the same mouse, the count was averaged before inclusion in the data. Whole images were processed in Adobe Photoshop, rotated where necessary, and adjusted so that all channels were equally visible.
Dissection and RNA extraction. The organs of Corti of four-day-old (P4) mice were dissected during a fixed time window (between 6 and 7.5 hours after lights on) to avoid circadian variation, and stored at -20°C in RNAlater stabilisation reagent (Ambion and sequenced on an Illumina HiSeq 2500 machine as paired-end 125bp reads. The resulting reads were quality checked using FastQC 0.11.4 (Andrews 2010) and trimmed with Trimmomatic 0.35 (Bolger et al. 2014); adapters were removed, trailing ends were clipped where the quality was low and a sliding window approach used to control for quality across the entire read. Finally, reads with 36 basepairs or fewer were discarded, because the shorter a read, the less likely it is to map uniquely to the genome. Reads were assembled to GRCm38 using Hisat2 version 2.0.2beta (Kim et al. 2015).
Bam files were soft-clipped beyond end-of-reference alignments and MAPQ scores set to 0 for unmapped reads using Picard 2.1.0 (Broad Institute, Cambridge, Mass. http://broadinstitute.github.io/picard) and checked for quality using QoRTS (Hartley and Mullikin 2015). The QoRTS tool also generates count data (in the same format as HTSeq), and these were used with edgeR (Robinson et al. 2010) to carry out a generalized linear model likelihood ratio test.
Transcriptome and network analysis. Sylamer 10-188 (van Dongen et al. 2008) was used to examine genes ranked in order of their misregulation from up-to downregulated for over-and underrepresented heptamers in their 3' UTRs. Ingenuity Pathway Analysis (IPA) was used to generate potential upstream regulators for the affected genes, using the causal network analysis module. The WGCNA R package (Langfelder and Horvath 2008) was used to carry out weighted gene correlation network analysis. Gene counts from the QoRTS tool were prepared for WGCNA using DESeq2 (Love et al. 2014) and transformed with a variance stabilising transformation. Batch effects were controlled for using the limma R package (Ritchie et al. 2015). Module enrichment analysis was carried out using PANTHER v14 (Mi et al. 2019), and oPOSSUM (Kwon et al. 2012) was used to assess transcription factor binding site overrepresentation in each module. Motif detection in a set of genes can be affected by differing GC composition in the genes used as a "background" set, so background gene sets of 4,000-5,000 genes were selected for each module such that the GC content matched that of the genes of interest. oPOSSUM assigns two scores to each transcription factor profile, the Z-score (which assesses whether the rate of occurrence of a given motif in the genes of interest differs significantly from the expected rate calculated from the background genes) and the Fisher score (which compares the proportion of genes of interest which contain a given motif to the proportion of the background genes containing that motif in order to determine the probability of a non-random association between the motif and the genes of interest) (Kwon et al. 2012). We chose to use a threshold of the mean + 1 standard deviation for each score.
RTPCR and qPCR. Organ of Corti RNA was normalised to the same concentration within each litter, then treated with DNAse 1 (Sigma, cat. no: AMPD1) before cDNA creation. cDNA was made using Superscript II Reverse Transcriptase (Invitrogen, cat. no: 11904-018)

or M-MLV Reverse
Transcriptase (Invitrogen, cat. no: 28025-013) or Precision Reverse Transcription Premix (PrimerDesign,. MicroRNA cDNA was made using the miRCURY LNA RT Kit (QIAGEN,cat. no: 339340). Primers for sequencing cDNA for testing differential splicing were designed using Primer3 (Untergasser et al. 2012) (Appendix Table S3). Sanger sequencing was carried out by Source Bioscience and analysed using Gap4 (Bonfield et al. 1995). Quantitative RT-PCR was carried out on a CFX Connect qPCR machine (Bio-Rad), using probes from Applied Biosystems and QIAGEN (see Appendix Table S3 for primer/probe details) and Sso-Advanced Master Mix (Bio-Rad, cat. no: 1725281)  are specifically expressed in hair cells, for which Jag1 was used as an internal control, because it is expressed in supporting cells of the organ of Corti (Morrison et al. 1999;Zine et al. 2000). For all other genes and microRNAs, the quantity of sensory tissue present was checked using Jag1, and pairs were only used if their Jag1 levels did not differ by more than ±20%. For the microRNA qPCR, the internal control was Mir99a, which is expressed in most cell types in the cochlea (Friedman et al. 2009). At least three technical replicates of each sample were carried out for each reaction, and at least four biological replicates were tested per probe (see legends of Appendix Figures S1, S11 for numbers for each probe).
Statistics. For qPCR and synapse counts, the Wilcoxon rank sum test was chosen to determine significance, because it is a suitable test for small sample sizes and populations of unknown characteristics (Bridge and Sawilowsky 1999). For repeated ABR threshold analyses, the thresholds were not normally distributed, so the data were first transformed using the arcsine transformation then analysed using separate linear models for each frequency with a compound symmetric covariance structure and restricted Maximum Likelihood Estimation (Duricki et al. 2016 Axioskop 2 microscope with a Plan Neofluar 63x 1.4NA objective was used to examine slides, and photos were taken using a Zeiss Axiocam camera and the associated Axiocam software. Images were processed in Adobe Photoshop; minimal adjustments were made, including rotation and resizing.
Where image settings were altered, the adjustment was applied equally to wildtype and mutant photos and to the whole image.
Prediction of potential causal regulatory networks. In our previous analysis (Lewis et al. 2016), we used interactions from the literature to connect miR-96 to as many of the misregulated genes in Mir96 Dmdo homozygotes as possible. In order to automate this procedure, a custom script (prediction of potential causal regulatory networks, PoPCoRN) was written to make use of publically available regulatory data from ArrayExpress and the associated Expression Atlas (Petryszak et al. 2016;Athar et al. 2019), ORegAnno (Lesurf et al. 2016), miRTarBase (Chou et al. 2018), TransmiR (Tong et al. 2019) and TRRUST (Han et al. 2018) (Table 3). All these data are based on experimental evidence, although we did make use of human regulatory interactions by converting human gene IDs to mouse gene IDs where there was a one-to-one orthologue (using Ensembl). Interactions from our previous network (Lewis et al. 2016), which were obtained from the literature using Ingenuity IPA, were also added, as were regulations reported in (Hertzano et al. 2007), which is not available through  the misregulated genes were then identified, and the direction of misregulation of intermediate genes predicted. Starting with the known misregulated genes, each upstream regulator was given a score based on the direction of regulation of the known misregulated genes (Appendix Figure S18).
This process iterated until the gene(s) at the top of the cascade (in this case, Mir96, Mir183 or Mir182) were reached. Consistent links were kept, the shortest paths between microRNA and misregulated genes identified, and the final network was written out in the simple interaction format (sif) and viewed using Cytoscape (Shannon et al. 2003) (Appendix Figure S18). Data and software availability: The datasets and computer code produced in this study are available in the following databases: • RNA-Seq data: ArrayExpress E-MTAB-5800 (https://www.ebi.ac.uk/arrayexpress/experiments/E-MTAB-5800).

Conflicts of interest:
The authors declare that they have no conflicts of interest.    . Mean ABR thresholds for wildtype and heterozygous Mir183/96 dko mice subjected to noise exposure (95dB SPL, 8-16kHz, for two hours), before exposure and 1, 3, 7, 14 and 28 days after. An increase in thresholds at 12kHz and above is seen 1 day after exposure but thresholds have returned to normal by 28 days after. Significant differences (P < 0.05, mixed linear model pairwise comparison) are marked by * in blue (for a significant difference between wildtype unexposed mice and heterozygous unexposed mice) or orange (for a significant difference between wildtype noiseexposed mice and heterozygous noise-exposed mice). Wave 1 amplitudes at 12kHz and 24kHz are shown for each time point. No obvious effect is visible at 12kHz, but at 24kHz, one day after noise exposure, wave 1 was too poorly defined to measure the amplitude in all heterozygotes and all but one wildtype. By 28 days after exposure, both wildtype and heterozygote amplitudes have recovered to the normal range. 12 wildtype mice and 12 heterozygote mice were tested; 6 wildtypes were noise-exposed (violet) with 6 unexposed controls (green), and 6 heterozygotes were noiseexposed (orange) with 6 unexposed controls (blue). Error bars are standard deviation. The grey area on the threshold plots indicates the octave band of noise (8-16kHz).    showing predicted upstream regulators which may be responsible for some of the misregulation . Misregulated genes are arranged on the lowest row, coloured according to nk/red = upregulated, green = downregulated in mutants). The top row(s) contain predicted regulators (orange = predicted upregulation, blue = predicted downregulation).
Predicted links inconsistent with the observed misregulation have been removed. The int the colour indicates the level of observed or predicted misregulation. Dotted lines represent indirect regulation, and solid lines direct regulation. The z-score of each network, which is both a prediction of the direction of misregulation of the root regulator and a measure of the match of observed and predicted gene misregulation, is shown in the figure. A significant z-score is one with an absolute value greater than 2. A negative score indicates downregulation and a positive score upregulation 96 is not one of the identified upstream regulators.
Mir183/96 dko RNA-seq data, showing predicted upstream regulators which may be responsible for some of the misregulation the lowest row, coloured according to nk/red = upregulated, green = downregulated in mutants). The top row(s) contain predicted regulators (orange = predicted upregulation, blue = predicted downregulation).
Predicted links inconsistent with the observed misregulation have been removed. The intensity of the colour indicates the level of observed or predicted misregulation. Dotted lines represent indirect score of each network, which is both a prediction root regulator and a measure of the match of observed and score is one with an absolute value greater than 2. A negative score indicates downregulation and a positive score upregulation of Figure 9. Diagram of the modes of action of mutant microRNAs in the outer hair cell. A) In a wildtype hair cell, miR-96 represses some genes completely (here represented by Zeb1 and Nr3c1) and