Abstract
Sensory hair cell loss is the major cause of hearing and balance disorders. Mammals are incapable of sustained hair cell regeneration, but lower vertebrates can regenerate these mechano-electrical transducers. We present the first comprehensive transcriptome (by mRNA-Seq) of hair cell regeneration in the chick utricle. We provide pathway and pattern annotations and correlate these with the phenotypic events that occur during regeneration. These patterns are surprisingly synchronous and highly punctuated. We show how these patterns are a new resource for identifying components of the hair cell transcriptome and identify 494 new putative hair-cell-specific genes and validate three of these (of three tested) by immunohistochemical staining. We describe many surprising new components and dynamic expression patterns, particularly within NOTCH signaling. For example, we show that HES7 is specifically expressed during utricle hair cell regeneration and closely parallels the expression of HES5. Likewise, the expression of ATOH1 is closely correlated with HEYL and the HLH inhibitory transcription factors ID1, ID2, and ID4. We investigate the correlation between fibroblast growth factor signaling and supporting cell proliferation and show that FGF20 inhibits supporting cell proliferation. We also present an analysis of 212 differentially expressed transcription factor genes in the regenerative time course that fall into nine distinct gene expression patterns, many of which correlate with phenotypic events during regeneration and represent attractive candidates for future analysis and manipulation of the regenerative program in sensory epithelia and other vertebrate neuroepithelia.
Introduction
Loss of sensory hair cells is the major cause of hearing and balance disorders (Marazita et al., 1993; Blanchfield et al., 2001). Sensory hair cells reside in sensory epithelia that transduce sound and movement. In mammals, once these epithelia mature, they are incapable of sustained regeneration (Forge et al., 1993; Warchol et al., 1993); however, lower vertebrates have a robust regenerative capacity in their sensory epithelia (Corwin and Cotanche, 1988; Ryals and Rubel, 1988; Warchol and Corwin, 1996). Avian hair cell death triggers one of two outcomes in the surrounding supporting cells. They can either convert directly into new hair cells (Roberson et al., 2004; Duncan et al., 2006), a process called phenotypic conversion, or they can reenter the cell cycle and generate new hair cells and supporting cells by regenerative proliferation (Girod et al., 1989; Stone and Cotanche, 1994). Within the vertebrate ear, both the vestibular (balance) and the cochlea (auditory) systems rely on hair cells to transduce mechanical movements into electrical stimuli. The vestibular system is morphologically conserved across diverse vertebrate species, whereas the mammalian auditory system has evolved specialized structures not present in other vertebrates (Meyers and Corwin, 2008).
Little is known about the genetic mechanisms that regulate regeneration in the sensory epithelia. One exception is the NOTCH pathway, which has been investigated in both mammals and birds during hair cell development and regeneration (Lindsell et al., 1996; Lanford et al., 1999; Stone and Rubel, 1999). Such studies demonstrated that the transcription factors ATOH1 and various HES genes play important roles in determining hair cell and supporting cell fates via reciprocal inhibitory loops. Notably, the upstream regulators of these transcription factors and the downstream mechanisms that further specify a hair cell are largely unknown.
Our group conducted the first large-scale gene expression analysis of the regenerative process in the avian inner ear specifically focused upon changes in transcription factor gene expression (Hawkins et al., 2007). Here, we present the first comprehensive transcriptome, by RNA-Seq, of hair cell regeneration in the chick utricle across a 7 d time course from the first stages of response to damage through to the production of new hair cells by regenerative proliferation. We provide a considerable amount of pathway and pattern annotation and correlate the gene expression data with the proliferation of supporting cells, the production of new hair cells by phenotypic conversion, and the later production of hair cells by regenerative proliferation. We also describe the major discernible patterns and pathways, some of which are surprising and dynamic and show how these are a new discovery resource for accurately identifying components of the hair cell transcriptome. Finally, we investigate the correlation between fibroblast growth factor (FGF) signaling and the control of supporting cell proliferation and present a clustering analysis of gene expression changes for 212 differentially expressed transcription factors in the regenerative time course, the vast majority of which have never been studied in regeneration and represent attractive candidates for future analysis and manipulation of the regenerative program in many vertebrate systems.
Materials and Methods
Chick utricle cultures and isolation of sensory epithelia.
Organotypic cultures of the chick utricle (extracted from both sexes) were prepared by previously described methods (Matsui et al., 2002). Utricles were treated with 1 mm streptomycin for 24 h. Untreated cultures were maintained in parallel and served as matched controls. After 24 h, all specimens were either harvested for analysis or were rinsed and maintained in culture for 1–7 d in streptomycin-free medium and fed at 2 d intervals. The pure sensory epithelia, consisting of only hair cells and supporting cells, were isolated from the underlying tissues either immediately after streptomycin treatment (0 h time point) or after 24–168 h of recovery in vitro. Utricles were rinsed and then incubated for 30 min in thermolysin (500 μg/ml at 37°C). Isolated epithelia for each time point and treatment group were pooled and dissolved in TRIzol reagent (Invitrogen). At least two biological samples (16 epithelia pooled from the streptomycin-treated group and 16 from the control group) were collected for every time point.
Immunohistochemistry.
Cultured utricles were fixed for 30 min with 4% paraformaldehyde, washed with PBS, incubated for 2 h in PBS with 5% normal horse serum, and incubated overnight in primary antibodies. Hair cells were labeled with antibodies against either otoferlin (HCS-1; a gift from Jeffrey Corwin University of Virginia, Charlottesville, VA) or Mysoin VIIa (1:500; Proteus). Antibodies for AP3B2, MDFI, and DNM1 were from Abcam. Specimens were rinsed with PBS and incubated for 2 h with secondary antibodies conjugated with fluorescent markers (Alexa Fluor 488 or 546; Invitrogen). Specimens were imaged with confocal microscopy (LSM 700; Zeiss) and processed with Volocity software (PerkinElmer).
Quantification of cell division and hair cell recovery.
Proliferation was assessed at 1–7 d after aminoglycoside antibiotic treatment. Cultures received BrdU (3 μg/ml) for the final 4 h in vitro. After fixation, utricles were processed for BrdU immunohistochemistry, an assay system that has been well described previously (Matsui et al., 2002), and imaged by confocal microscopy. BrdU-labeled cells were quantified from six 100 × 100 μm regions within the extrastriolar portion of each specimen. Data were obtained from 6–10 utricles per time point per treatment group. Hair cell recovery was quantified using similar methods. Regenerated hair cells were labeled with an antibody against myosin VIIa (see Immunohistochemistry, above). Confocal images were obtained and hair cells were quantified in six 100 × 100 μm extrastriolar regions per utricle.
FGF signaling and supporting cell proliferation.
Cultured utricles were treated with either FGF20 (to activate FGFR1 and/or FGFR3) or SU5402 (to inhibit FGFRs). Utricles were treated for 24 h with 1 mm streptomycin, as described in Utricle culturing, above. Cultures were rinsed and given Medium-199 supplemented with 1% FBS and containing either FGF20 (1 μg/ml with 1 μg/ml heparin sulfate; R&D Systems) or SU5402 (30 μm; Santa Cruz Biotechnology). Control cultures received either 1 μg/ml heparin sulfate (FGF20 experiments) or 0.1% DMSO (SU5402 experiments). Each treatment group contained 10–12 specimens. Utricles were incubated in media for 48 h and proliferation was measured as in Quantification of cell division, above. Statistical significance was assessed using Student's t test.
RNA-Seq preparation.
Samples from each time point were processed using Illumina mRNA-seq or TrueSeq preparation kits. In brief, mRNA was selected by oligo-dT magnetic beads from 1 μg of total RNA and fragmented. First-strand cDNA was generated using random primers. Second-strand synthesis, end repair, addition of a single A base, and adaptor ligation were then performed. Each RNA-seq library was DNA sequenced using either the Illumina Genome Analyzer IIx or HiSeq 2000. In all cases, biological replicate samples from pure sensory epithelia were analyzed. The average correlation coefficient between biological replicates was 0.9423. In some cases, we also ran technical replicates. The average correlation coefficient between technical replicates was 0.9979.
RNA-Seq data analyses.
Raw reads in FASTQ format were aligned to the Ensembl Gallus gallus reference genome (WUGSC2.1 E66) using Tophat v1.4.0 (Trapnell et al., 2009). The output.bam file was processed by Partek Genomics Suite version 6.6 to assemble reads into transcripts and estimate their abundances. A gene model dataset combining Ensembl and National Center for Biotechnology Information annotations was used to generate reads per kilobase of exon per million fragments mapped (RPKMs) for known genes and potential novel transcripts. Statistical significance levels were calculated by one-way ANOVA. Biological replicates had an average R of 94%. Reads mapped to different isoforms of a given gene were combined together for analysis. At least one sample across the entire time course was required to be ≥0.5 RPKM. At least one time point had to have a significant fold change (p < 0.05 and fold change ≥1.8). The same filters were applied to individual time points to produce a differentially expressed dataset. Gene lists from clustering analysis were uploaded to g:Profiler and analyzed by gene group functional profiling. Pathways that were predicted from g:Profilers database of literature-supported interaction datasets were visualized by Cytoscape. All sequence data and annotations have been submitted to NCBI Gene Expression Omnibus (GEO) and are available online.
Quantitative real-time PCR.
Total RNA was isolated from a time course of treated and control chicken sensory epithelia at all time points and was converted to cDNA. Quantitative real-time PCR (qRT-PCR) was performed using the ABI TaqMan Gene Expression Master Mix (Applied Biosystems) on an ABI 7900HT real-time PCR machine. All of the qRT-PCR assays were obtained from Integrated DNA Technologies and were run in triplicate. The expression of 18S rRNA was used as a normalization control. Twenty genes (including ATOH1 and HES5) were selected for validation across every time point (in triplicate) across a new parallel set of biological samples (a total of 660 qRT-PCR assays). The average correlation between the fold change in qRT-PCR and RNA-seq was 86% and in all cases the trends agreed.
Results
We used an in vitro organotypic culture system in which hair cell regeneration is similar to that observed in vivo (Warchol, 1999, Matsui et al., 2002). Whole utricles were placed in culture and subjected to 24 h of aminoglycoside antibiotic treatment, which kills ∼93% of the hair cell population (Warchol and Montcouquiol, 2010). Untreated utricles were cultured in parallel and served as controls. After aminoglycoside antibiotic treatment, the sensory epithelia of some specimens (along with untreated controls) were removed for RNA-Seq; this is the zero time point (see Materials and Methods). Other specimens were rinsed and cultured for 1–7 d. We sampled at 24 h intervals up to 168 h. All of our RNA-Seq data were derived from pure sensory epithelia without any detectable stromal contamination. These sensory epithelia consist of hair cells and supporting cells, the latter of which act as progenitors for hair cell regeneration. The ratio of supporting cells to hair cells is ∼4:1 (Goodyear et al., 1999). The initial hair cell density in these cultures is ∼180/10,000 μm2 and drops to ∼12/10,000 μm2 after antibiotic treatment. Large numbers of the supporting cells undergo proliferation within 96 h after antibiotic treatment, resulting in the eventual replenishment of hair cell numbers. In addition, a minority of new hair cells are produced by direct phenotypic conversion within the first 2–3 d after antibiotic injury. It became clear from our initial samples that surprising changes in gene expression were occurring in the 48–72 h window of regeneration. We therefore sampled at 6 h intervals across the 48–72 h window (56, 60, and 66 h). In all cases, at least two biological replicates were analyzed at every time point. The Materials and Methods section include information on data reproducibility. We conducted qRT-PCR on RNA samples from a separate time course across all 11 time points on 20 genes to validate the mRNA-Seq data. In all cases, these trends agreed (data are available online).
Defining the regenerative transcriptome
We set our threshold for differentially expressed genes at ≥1.8-fold and a p-value <0.05 at any one time point for treated versus control cultures. Raw sequences have been submitted to NCBI GEO (accession number applied for) and all annotations are available online. We also required that the abundance level of each transcript was ≥0.5 RPKM (Mortazavi et al., 2008) in any sample. This resulted in a list of 3661 genes differentially expressed at any single time point.
Correlations between expression patterns and functional annotations
Lists of upregulated and downregulated genes at individual time points were uploaded to the DAVID Bioinformatics Resource 6.7 (da Huang et al., 2009) to perform gene ontology (GO) enrichment analysis (data are available online in supplemental Table 9). At early time points (unsurprisingly), sensory perception of sound processes are depleted (∼10% of genes with p-values <10−5), reflecting the death of hair cells. For upregulated genes, the GO terms are enriched for cell cycle annotations from 24 to 120 h (∼20% of genes and p-values <10−5). This switch in cell cycle (on at 24 h and off at 120 h) agrees with studies that observed mitotically regenerated hair cells ∼6 d after aminoglycoside antibiotic treatment (Cotanche et al., 1994). After 120 h, the number of differentially expressed genes decreases by 52%, reflecting the cessation of the proliferative program. At these later stages, GO annotations point to switches in extracellular matrix production and glycoproteins. This coincides with the period when mature hair cells develop from regenerative proliferation of the supporting cells.
To identify more specific patterns, we clustered all differentially expressed genes by unsupervised clustering within self-organizing maps (SOMs; Kohonen, 1982) to identify 36 patterns across the time course (Fig. 1). This number of distinct patterns was arrived at empirically by deriving smaller (20 centroids) and increasingly larger numbers of unsupervised clusters. It is a compromise between deriving duplicate patterns (e.g., cell cycle patterns) and observing discrete interesting patterns that are enriched for different cellular processes. For example, centroids 19, 25, and 31 in Figure 1 are highly enriched for cell cycle programs, indicating that this initiates at 24 h and has largely ceased by 120 h. We also conducted functional annotations and network analyses on these SOM clusters using the DAVID, g:Profiler (Reimand et al., 2011), and Cytoscape tools (Smoot et al., 2011). Several of these functional correlations are shown in Figure 1.
Differentially expressed gene clusters analyzed by SOMs. Differentially expressed genes were clustered into 36 centroids by SOMs (Partek). Each x-axis indicates individual time points. The y-axis shows the average normalized fold change (log2) for all the genes in the centroid, with the bars showing the range. The number within each centroid refers to a gene cluster referred to in the text. The specific number of genes and gene IDs in each cluster are listed in supplemental Table 4. Some of the correlations between DAVID functional annotations and patterns are shown within each centroid, with the following codes: AAB, amino acid biosynthesis; AAM, amino acid metabolism; AP, apoptosis; CA, channel activity; CB, calcium binding; CC, cell cycle; CM, cell migration; CYT, cytoskeleton; DD, DNA damage; EM, extracellular matrix; ER, endoplasmic reticulum; GLY, glycosylation; HLH, helix loop helix genes; IFN, interferon responses; MIB, metal ion binding; MP, membrane proteins; NB, nucleolus biosynthesis; ND, neural development; NOS, nuclear organelle synthesis; PAT, patterning; RB, RNA binding; RD, regulation of differentiation; SPS, sensory perception of sound; SR, stress response; TM, transmembrane transport; and VGC, voltage gated channels.
Correlating the time course with regenerative phenotypes
Studies of hair cell regeneration have generally used two phenotypic measurements: (1) supporting cell proliferation measured by incorporation of BrdU into DNA (Gratzner, 1982) and (2) the production of mature hair cells measured by the presence of hair-cell-specific proteins such as myosin VIIA (Avraham et al., 1995). We quantified these across the regenerative time course. Figure 2 shows the proliferation index and hair cell recovery values. DNA replication (Fig. 2A) is a highly punctuated process, with a peak at 48 h followed by a steep decline. This peak contrasts with the spectrum of cell-cycle-related transcripts (Fig. 1, top), which number >300 different GO-annotated gene products and collectively show a long plateau of expression. These transcripts only decline in expression as a group at 96–120 h, presumably because the replication machinery is used to ensure complete cytokinesis.
Phenotypic measurements across the regenerative time course. A, DNA replication in supporting cells was measured by a BrdU incorporation assay. The x-axis indicates individual time points. The y-axis shows the mean number of BrdU-labeled cells per 10,000 μm2 ± SD in control and streptomycin-treated organotypic cultures. B, The number of mature hair cells was quantified by measuring MYO7a-labeled cells. Chicken utricle sensory epithelia organotypic cultures were treated with streptomycin and immunohistochemically labeled for MYO7a to detect mature hair cells. The x-axis indicates individual time points. The y-axis shows the mean number of hair cells per 10,000 μm2 ± SD.
Measuring the appearance of new hair cells by MYO7a staining is complicated by the persistence of dead (but unextruded) hair cells within the injured sensory epithelia. There is nevertheless evidence from MYO7a staining (Fig. 2B) that hair cell numbers decline through the 48 h time point and are then replenished by 72 h (consistent with a period of phenotypic conversion in that time frame), followed by an ∼2-fold increase by 168 h (from regenerative proliferation).
Identifying components of the hair cell transcriptome
A more refined measure of hair cell production is provided by tracking a panel of 32 genes that are known to mark hair cells within the utricle sensory epithelia (these include MYO7a) in the RNA-Seq data shown in Figure 3. All of these showed a >2-fold drop (range of −2-fold to −16-fold) in gene expression after aminoglycoside antibiotic treatment, reflecting hair cell loss, and then showed various patterns of recovery, with all showing higher expression levels at 168 h. Interestingly, many of them exhibited either a peak or plateau of expression at ∼72 h, coincident with the probable onset of phenotypic conversion and the production of a minority of new hair cells. The majority of these markers do not return to wild-type levels by 168 h, which is consistent with the replenishment of ∼25% of normal hair cell numbers by this time.
Expression profiles of 32 known hair cell markers. The x-axis indicates individual time points. The y-axis shows the fold change of streptomycin-treated versus control samples on a log2 scale. These sentinel genes are significantly downregulated at the 0 and 24 h time points and showed a recovery trend toward the end of the time course.
We used this pattern as a filter to identify new hair-cell-specific markers. Defining the hair cell transcriptome remains a challenge. At present, only a small proportion of the hair cell transcriptome has been identified. For example, various array-based methods (McDermott et al., 2007; Hertzano et al., 2011; Sinkkonen et al., 2011) and proteomic studies (Herget et al., 2013; Shin et al., 2013) have identified a few hundred hair cell markers. Therefore, we implemented a filter on the entire list of 3661 genes that required a >2-fold expression drop at 0–24 h and a relative recovery by 168 h. This resulted in a list of 526 genes (some examples are shown in Table 1), within which were our 32 hair cell “sentinel” genes. We then derived 25 SOM clusters from this set (Fig. 4). All of these clusters showed the same trends at the beginning and end of the time course, but there were reproducible fluctuations at 48–72 h, which is consistent with phenotypic conversion of supporting cells to hair cells. This interpretation is supported by previous studies showing that phenotypic conversion occurs at ∼3 d after antibiotic removal (Roberson et al., 2004) and by our measurements of MYO7a protein.
Examples of putative hair-cell-specific genes identified in this study
SOMs of potential hair cell markers. Genes that passed the filters for potential hair cell markers (downregulated >2-fold at 0–24 h and a recovery trend by the 168 h time point) were clustered into 25 groups by SOMs (Partek). The x-axis of each group indicates individual time points. The y-axis shows the normalized fold change.
To determine whether our filtered set was indeed enriched for bona fide hair-cell-specific genes, we chose three with available antibodies for validation by immunohistochemistry. The first of these, DNM1, encodes a member of the dynamin subfamily of GTP-binding proteins. It has been implicated in signal transduction and vesicle trafficking (Sever, 2002; Wiejak and Wyroba, 2002). DNM1 colocalized with MYO7a within all hair cells (Fig. 5A), suggesting that it may be a part of the stereocilia bundle. The second gene, AP3B2, encodes a subunit of the adapter protein complex 3 and plays a role in neurotransmitter release (Newman et al., 1995; Faúndez et al., 1998). AP3B2 was specifically expressed in the hair cell cytoplasm (Fig. 5B) and at similar levels across all hair cells. The third gene was MDFI, which encodes a MyoD family inhibitor. This transcription factor negatively regulates other myogenic family proteins, interacts with Axin, and regulates the JNK and WNT pathways (Kusano and Raab-Traub, 2002). Approximately 90% of utricle hair cells were labeled for MDFI (Fig. 5C), but the intensity of nuclear staining varied, with ∼33% being intensely stained.
Immunohistochemical staining of novel hair cell markers. Chicken utricles were labeled with antibodies to known hair markers (MYO7a or otoferlin) and to potential novel hair cell markers identified in the present study. Scale bar, 16 μm. A, DNM1 (green) and MYO7a (red) are colocalized within all hair cells and are expressed at similar levels. B, Hair cells are labeled for otoferlin (green). AP3B2 (red) is expressed homogeneously within the cytoplasm of hair cells. C, MYO7a (green) labels hair cells and MDFI (red) labels the nuclei of a subpopulation of hair cells at various intensity levels. The images show the apical surface of the sensory epithelium with supporting cells filling the gaps between the labeled hair cells.
We conclude that the 494 filtered genes are enriched for new hair cell markers. Furthermore, in addition to MDFI, an additional 13 transcription factor genes are present in our putative hair cell clusters. These are noteworthy because so few hair-cell-specific transcription factors have been previously described. Of the 14 transcription factor genes listed in Table 2, three are known hair-cell-specific transcription factors (POU4F3, GFI1 and LHX3), one is the MDFI gene we validated, and the remaining 10 are new candidates for further investigation of hair cell gene regulation. It is worth noting that our new putative hair cell markers, unsurprisingly, have little overlap with the few hundred described in the various array-based and proteomic-based studies cited above.
Examples of significant transcription factor gene clusters
Network analysis of the entire 526 putative hair-cell-specific genes provides independent evidence for a proportion of them interacting. Figure 6 shows one part of this overall g:Profiler (Reimand et al., 2011)-predicted hair cell network in which 94 of the detectable genes in the regenerative transcriptome are linked within a network (supported by literature citations to be found within the g:Profiler database) that includes one known hair cell marker (KCNA2), one of our new validated markers (DNM1), and 13 of our putative new hair-cell-specific markers. Therefore, in other cell systems there is independent evidence that these gene products directly interact.
Examples of gene networks within the clustered data. A, Network of interactions between putative hair-cell-specific genes. All 526 hair-cell-specific gene names from the clustering analysis were uploaded to g:Profiler and analyzed as described in the Materials and Methods. This figure includes only genes detectably expressed in the regenerative time course. Red circles denote genes within the 526 putative hair-cell-specific input set. Red lines indicate interactions between those genes/gene products. Blue circles denote additional genes within the differentially expressed subset. Black circles denote genes that are detectable but not differentially expressed. Black lines denote interactions between genes/gene products. B, Network of interactions between some of the transcription factor genes within clusters 13 and 15 (Fig. 8) exhibiting a dip in expression at 120 h. This includes only genes that are detectably expressed in the regenerative time course. Color coding as in A except red genes are differentially expressed transcription factors in these clusters.
Components of known pathways and processes within the differentially expressed genes
We next extracted components of known and interesting signaling pathways/gene families from the overall dataset. We extracted data for 47 pathways/gene families that have been implicated in hair cell development or regeneration (data are available online). Here, as examples of the rich information content in these data, we highlight observations on just two of these pathways that play key roles in sensory epithelia biology: NOTCH (Lanford et al., 1999; Stone and Rubel, 1999) and FGF (Alvarez et al., 2003) signaling.
NOTCH pathway
Our data include significant alterations in the expression of 25 components of NOTCH signaling. One of these is ATOH1. Elevated levels of ATOH1 are associated with hair cell differentiation and initiate a process of lateral inhibition in which ATOH1 represses HES genes. Elevated ATOH1 also leads to enhanced levels of Delta and Jagged. These bind to NOTCH receptors on adjoining cells. Processing of NOTCH then leads to the expression of various HES transcription factor genes that repress ATOH1 expression within these neighboring cells and direct them toward a supporting cell fate (Morrison et al., 1999; Zine et al., 2001). Previous studies have suggested that the major HES paralogs downstream of NOTCH signaling in the utricle are HES1 and HES5 (Heitzler et al., 1996; Lanford et al., 2000). HES1 does not show significant changes in gene expression in our dataset, although it is detectably expressed. However, HES5 is abundantly expressed and dynamically regulated during avian utricle regeneration. Notably, the expression of HES5 was sharply reduced immediately after ototoxic injury. This observation, combined with high levels of NOTCH expression, suggests that NOTCH signaling is tonically active in the normal utricle and that this is interrupted by hair cell death. We also noted a dip in HES5 expression in the 24–72 h interval in our original time course (Fig. 7A). Therefore, we investigated this in more detail at 6 h intervals (54, 60, and 66 h; Fig. 7B). HES5 expression dramatically increases across the time course and appears to do so in an oscillatory manner. Some HES genes are known to act within natural oscillatory networks (Bessho et al., 2001; Hirata et al., 2002). HES5 expression appears to drop within the 48–72 h interval. This corresponds to the time frame for phenotypic conversion and correlates well with the intermediate peaks in hair cell marker expression. A surprising observation is the closely parallel pattern of gene expression exhibited by the HES7 gene (Fig. 7B). This gene plays a key oscillatory role in somitogenesis (Niwa et al., 2011), but has never been investigated or implicated in inner ear development. It is expressed at an ∼10-fold lower level than HES5 in our dataset, but it also exhibits a parallel dip in expression at 54 h and a >8-fold increase in expression by the end of the regenerative time course.
Changes in specific components of the NOTCH and FGF signaling pathways during regeneration. The x-axis indicates individual time points. The y-axis shows the fold change of streptomycin-treated versus control samples on a log2 scale. A, Expression profiles of ATOH1, DLL1, and HES5 at 24 h intervals across the 168 h time course. B, Expression profiles of HES5 and HES7 with the addition of detailed sampling within the 48–72 h period at 6 h intervals. C, Expression profiles of ATOH1, HES6, and HEYL. D, Expression profiles of differentially expressed intracrine FGFs (FGF12 and FGF14). E, Expression profiles of differentially expressed paracrine FGFs (FGF3, FGF16, and FGF20). F, Expression profiles of FGF19, FGFR1, and FGFR3.
ATOH1, like HES5, is abundant within the regenerative transcriptome. It exhibits an ∼2-fold increase in abundance over the time course with dips at 54 and 120 h (Fig. 7C). The latter is coincident with the time period when cell division is ceasing and hair cell differentiation is occurring. This pattern of ATOH1 expression is highly correlated with expression of the HEYL gene (hairy/enhancer-of-split related with YRPW motif-like). This coexpression has not been previously described and our observation suggests that a possible upstream coregulator of these two bHLH genes may exist within the regenerating utricle sensory epithelia. One downstream target of ATOH1 is HES6 (Qian et al., 2006), which exhibits much larger fold changes than ATOH1, reaching a 10-fold increase in expression relative to the controls at 72 h (Fig. 7C). It is interesting that the drop in HES5 expression at 54 h does not correlate with any corresponding major increase in ATOH1 expression (a known repressor of HES5). Therefore, some effector(s) other than ATOH1 may downregulate HES5 expression at this time point.
Together, these observations highlight several coordinate patterns of gene expression that involve previously unexplored components of NOTCH signaling. They also point to a highly punctuated switch occurring in NOTCH signaling in the ∼48–72 h window of regeneration consistent with phenotypic conversion.
FGF signaling
The FGF family and their receptors are key regulators of development in the inner ear (Schimmang, 2007). Twelve members of the FGF gene family are detectably expressed in our dataset. Of these, six (FGF3, FGF12, FGF14, FGF16, FGF19, and FGF20) are differentially expressed. FGF3 and FGF20 mouse gene knock-outs result in inner ear malformations (Alvarez et al., 2003; Hayashi et al., 2008), but FGF12, FGF14, FGF16, and FGF19 have not yet been investigated in hair cell development or regeneration. Of the five described FGF receptors, we can detect four (FGFR1, FGFR2, FGFR3, and FGFRL1) at above threshold levels, but only FGFR3 passes our filters. In Figure 7, the six FGFs that show large changes in gene expression are shown according to their known modes of action (Itoh and Ornitz, 2008). Two (FGF12 and FGF14) are intracrine FGFs that appear to operate as intracellular signaling molecules. They may have some functional redundancy, but in this case their gene expression patterns are quite different. FGF12 is initially downregulated (Fig. 7D) and recovers by 168 h. A more dramatic pattern is seen for FGF14, which is initially upregulated by 4-fold. It decreases in abundance, remaining downregulated throughout the remainder of the time course. There is a well established connection between FGF14 and the regulation of voltage-gated sodium (NAv) channel function (Lou et al., 2005), suggesting that the early stages of the regenerative time course may be a time of rapidly changing channel function in the supporting cells.
The FGF3, FGF16, and FGF20 genes encode classic paracrine FGFs that act through FGFRs. Of these, FGF16 and FGF20 are most closely related (Itoh and Ornitz, 2008). These gene expression patterns are shown in Figure 7E and some similarities can be seen. They both exhibit fluctuations in an overall upward direction, diverging at the end of the time course. In contrast, FGF3 is downregulated throughout the regenerative time course and only rebounds at the onset of hair cell differentiation in the 144–168 h time period. The final FGF that passes our filters is FGF19. This is a member of the small family of FGFs that have the ability to act at a distance in an endocrine manner (Itoh and Ornitz, 2008). This gene is dramatically downregulated between 0 and 54 h (Fig. 7F) and then exhibits a significant rebound in expression at 66–72 h, followed by an 8-fold drop in expression at 96 h and a rebound again at 144 h. This interesting pattern of gene expression parallels the two phases of hair cell production and suggests that this FGF may be an interesting candidate for functional follow-up.
The final components shown in Figure 7F are FGFR1 and FGFR3. FGFR1 did not pass our filters, but only narrowly. It is ∼10-fold higher in transcript abundance than FGFR3 and displays an expression pattern that is almost the reciprocal of FGFR3, starting 2-fold higher than controls and ending slightly lower, whereas FGFR3 starts 2-fold lower and ends the time course 2-fold higher than controls.
Although complex, the expression patterns of FGFs and FGFRs described here suggest a functional role for FGF signaling in utricular regeneration. Prior studies have shown that FGF20 interacts preferentially with FGFR3 (Itoh and Ornitz, 2011). Our data indicate that both FGF20 and FGFR3 transcripts are low shortly after ototoxic injury (when proliferation levels are high), but then increase later when regenerative proliferation stops and hair cell differentiation occurs. One interpretation of these findings is that FGF20 signaling may inhibit (rather than promote) supporting cell proliferation. To test this, we treated cultured utricles (n = 11) with exogenous FGF20 immediately after ototoxic injury. Cultures were maintained for 48 h and a proliferation index was measured by BrdU incorporation. Control cultures did not receive FGF20. Results indicated that FGF20 significantly reduced supporting cell proliferation (32.5 ± 9.9 BrdU-labeled cells/10,000 μm2 in control specimens vs 12.3 ± 9.6 labeled cells/10,000 μm2 in FGF20-treated specimens, p < 0.0001). In contrast, treatment for 48 h with an inhibitor of FGF receptors (SU5402) had no significant impact on proliferation (33.7 ± 13.2 labeled cells/10,000 μm2 in SU5402-treated cultures vs 26.8 ± 6.7 labeled cells/10,000 μm2 in control cultures, p = 0.21). Notably, application of FGF2 (the expression levels of which did not show significant changes in the regenerating utricle) can also inhibit supporting cell proliferation (Oesterle et al., 2000). Together, these data suggest that FGF signaling (specifically FGF20) may be a negative regulator of regenerative proliferation in the avian utricle.
Differentially expressed transcription factor genes
Among the differentially expressed genes, we were interested in identifying those encoding transcription factors that are potential key regulators of the regenerative program. We compiled a nonredundant list of transcription factors from three databases and recent publications (Thomas et al., 2003; Messina et al., 2004; Fulton et al., 2009; Vaquerizas et al., 2009; Ravasi et al., 2010; Zhang et al., 2012) and compared that with our list of differentially expressed genes. In this way, we identified 212 differentially expressed transcription factors (several of which are highlighted in Table 2). The vast majority of these have never been investigated in hair cell regeneration or in the development of the inner ear.
We next derived 16 SOMs from these gene expression profiles to identify distinct expression patterns. These are shown in Figure 8 and fall into nine patterns, with one additional group of four “complex” centroids (Fig. 8J).
Gene expression profiles of differentially expressed transcription factors clustered into 16 centroids by SOMs. The 212 differentially expressed transcription factors were grouped by SOMs (Partek) according to their expression patterns. The x-axis of each group indicates individual time points. The y-axis shows the normalized fold change on a log2 scale. Centroids were then manually grouped according to their shared overall expression patterns. See Differentially expressed transcription factor genes in the Results for a description of patterns A–J.
Figure 8A shows a centroid enriched for hair-cell-specific transcription factors. These comprise 21 genes and include known hair-cell-specific transcription factor genes (POU4F3, GFI1, and LHX3) and others identified in this study (ANKRD5, DPF3, NFATC1, and SKOR1). The hair cell expression filters that we imposed above on the total gene set were quite conservative. Therefore, the additional 15 transcription factors within this group may represent additional hair-cell-specific transcription factors. Figure 8C is enriched for genes that parallel the HES5 and HES7 patterns of gene expression with changes in expression coincident with phenotypic conversion listed in Table 2. Particularly interesting in this cluster is the BTG1 gene, which appears to have antiproliferative and prodifferentiation functions (Rouault et al., 1992; Rodier et al., 1999), the interferon regulatory factor 4 (IRF4) gene (Grossman et al., 1996), and the Cbp/p300-interacting transactivator (CITED4; Bragança et al., 2002), none of which have been previously investigated or implicated in sensory epithelia regeneration. Figure 8I shows eight genes that exhibit peaks of expression in the 54–72 h window of the time course. These are attractive candidates for playing roles in the phenotypic conversion process. Among the most highly expressed in this class are two implicated in the NOTCH pathway: MAMLD1 (mastermind-like, a coactivator in NOTCH signaling) and RBPJL (similar to the RBPJ gene in NOTCH signaling; Table 2). Figure 8G shows clusters that share a pronounced drop at 120 h coincident with cell cycle cessation. Seven of these 25 genes, including NOTCH1, show evidence for network interactions and tie together a predicted network of 138 genes that are detectable in the time course (Fig. 6B). This network provides strong confirmatory evidence that these genes are functionally connected and provides insights into their possible downstream effectors. The genes shown in Figure 8G are also of particular note because they contain seven of the 20 helix loop helix (HLH) or basic helix loop helix (bHLH) family of genes present in the list of differentially expressed transcription factors. Of these, five (ATOH1, HEYL, ID4, ID2, and ID1) are highly correlated and are particularly interesting. The ID gene family consists of four HLH paralogs (ID1–ID4). They encode dominant-negative inhibitors of bHLH genes such as ATOH1. ID1, ID4, and ID2 all show a strong positive correlation with ATOH1 expression during utricle sensory epithelia regeneration. A preliminary computational analysis of the promoter-proximal regions of these genes did not reveal any shared DNA sequence motifs (data not shown). Nevertheless, coordinate regulation through more distal regulatory elements cannot be discounted. ID genes are among the most highly expressed transcription factor genes in the regenerating sensory epithelia. These observations suggest that ID proteins are present at very high levels and are probably in vast excess compared with many of their targets (activating bHLH proteins). It is clear that the pattern of ATOH1 and ID gene expression in the regenerating avian utricle sensory epithelia is highly correlated and thus very different from that described for the mouse auditory sensory epithelia (Jones et al., 2006), in which these genes appear to show reciprocal patterns of expression. These observations suggest that different regulatory circuits may operate in the auditory and vestibular epithelia to modulate the activity of ATOH1 and additional bHLH genes.
Discussion
Here, we have derived the most comprehensive description to date of the mRNA transcriptome during hair cell regeneration. Our previous study of this process (Hawkins et al., 2007) focused only on the first 48 h of regeneration and solely upon transcription factor genes (many of which at that time had no clear chicken orthologs). A total of 87 transcription factor genes identified in those earlier expression profiles overlap with the present study. Clearly, our new dataset represents a greatly expanded, much more complete, accurate, and highly complex genetic toolbox for investigations of sensory regeneration and for comparisons with the nonregenerative mammalian inner ear. In particular, in the present study, we have emphasized patterns and pathways of gene expression that correlate well with three phenotypes of the regenerative sensory epithelia; DNA replication/cell cycle control, phenotypic conversion, and regenerative proliferation. We also identified additional patterns that we predict are enriched for genes affecting processes such as apoptosis, RNA processing, and components of the hair cell transcriptome (Fig. 1). In the latter case, we tested our prediction directly by determining the hair cell specificity of three randomly chosen genes in this subset by immunohistochemistry. As predicted, all of these showed hair cell specificity.
A surprising observation from our dataset is the highly punctuated nature of the events that occur across this time course. DNA replication and the expression of cell cycle components appear quite synchronous and provide patterns of gene expression that are highly diagnostic. Likewise, the process of phenotypic conversion appears to occur in the 48–72 h time period and is likely to be reflected in highly dynamic patterns of expression for many genes in this time window, particularly many of the known components of Notch signaling. The high levels of proliferation observed in the utricle suggest that most new hair cells are created by cell division, rather than through phenotypic conversion. Therefore, signals that are entirely specific to the new hair cell population in this time window appear minor when this dilution effect is not taken into account. With that in mind, if the overall 2-fold increase in ATOH1 expression from 60 h through 72 h is hair cell specific, then this would reflect a very large induction of this transcription factor. In that same context, the changes in HES5 (considered to be a supporting cell marker), if they are occurring in all of the supporting cells in this same time frame, would be relatively smaller compared with ATOH1. It remains to be seen whether the additional dynamically expressed genes in this time interval mark supporting cells, new hair cells, or cell subpopulations.
We anticipate that this resource will prove immediately useful as a source of components to test systematically by knock-downs or overexpression analyses (Alvarado et al., 2011). We extracted considerable pathway and network annotations to enhance data accessibility and utility. Several of these pathways have been implicated previously in inner ear sensory epithelia biology. However, our annotations now identify the exact individual components within these pathways, their relative transcript abundances, and changes in gene expression. In some cases, this reveals surprising pathways and genes that have not been implicated previously in sensory epithelia function or regeneration. As noted in Notch pathways in the Results, above, several surprising components of Notch signaling (e.g., HES7) are present in our dataset or exhibit interesting patterns of coordinate expression (e.g., HEYL and ATOH1). In addition to the many transcription factors we have identified that have never been implicated in inner ear development or regeneration, there are numerous other surprising new candidates. Some notable examples occur among members of what appears to be the regenerative “secretome.” Prominent among these is leukemia inhibitory factor (LIF), a cytokine capable of inducing differentiation in numerous cell types, including leukemic and neuronal cells (Chen et al., 2010; Mathieu et al., 2012). LIF transcripts, along with several additional components of the LIF signaling pathway (Ip et al., 1992), are abundant and differentially expressed during sensory epithelia regeneration. The regenerating pure avian sensory epithelium also produces a surprisingly complex mixture of what have been historically regarded as growth and differentiation factors of the hematopoietic system. These inflammatory modulators have not yet been investigated in inner ear sensory epithelia. For example, we observed robust expression of the chemokine CXCL14, which is found in zebrafish lateral line neuromasts (Long et al., 2000). Although the function of CXCL14 is not known, it is expressed in the chick utricle throughout the regenerative time course and is among the most abundant transcripts. Additional chemokines, cytokines, and their receptors are also differentially expressed in our dataset, including CCL17, CCL4, CXCL12, and CXCR4. In addition, we observed 14 interleukins or interleukin receptors within our differentially expressed dataset. Most prominent among these are IL18, IL16, IL22RA1, and IL6. Of these, IL6 is produced by many different epithelia and plays a role in epithelial-stromal interactions (Elner et al., 1992; Notara et al., 2010). All of these are attractive candidates for functional follow-up. Finally, our observations suggest novel roles for FGF signaling in the regenerating utricle. We observed robust expression of the intracrine factors FGF12 and FGF14, the function of which has not been investigated previously in the ear. In addition, our data suggest that FGF20, which regulates the differentiation of outer hair cells in the developing cochleae of mammals (Huh et al., 2012), may act to suppress regenerative proliferation in the chick utricle. Notably, enhanced expression of FGF20 begins at ∼3–4 d after ototoxic injury as regenerative proliferation is terminating and replacement hair cells are beginning to differentiate. It is possible that expression of FGF20 serves as a “stop” signal in the proliferative process.
One motivation for conducting this study was to identify the best candidates for controlling the regenerative cascade in birds and possibly affecting regeneration in the mammalian inner ear. In particular, the transcription factor genes we have highlighted here should provide useful starting points in deciphering the exact genetic circuitry of hair cell regeneration. This resource may also prove useful as a source of candidate genes for studies of other sustentacular stem cell populations that give rise to neurons and sensory cells in many nonmammalian vertebrates.
Notes
Supplemental material for this article is available at https://workspace.imperial.ac.uk/medicine/Public/PWPs/HimaAnbunathan/Ku.et.al.Archive.zip. Ten tables describing qRT-PCR, complete dataset, clustering, and pathway analysis. This material has not been peer reviewed.
Footnotes
This work was supported by the National Institutes of Health (Grants RC1DC010677 to M.L. and R01DC006283 to M.E.W. and National Institute on Deafness and Other Communication Disorders Grant P30DC04665).
The authors declare no competing financial interests.
- Correspondence should be addressed to Professor Michael Lovett, NHLI, Imperial College, Guy Scadding Building, Dovehouse Street, London SW3 6LY. m.lovett{at}imperial.ac.uk