Background
In humans, hearing loss is a potentially debilitating condition that affects more than 1.23 billion people worldwide and constitutes one of the world’s top ten causes of years lived with disability [
1]. The most common form of hearing loss, which represents 90% of all cases, is related to the degenerative effects of aging on hearing, i.e., age-related hearing loss or presbycusis. This is followed by inflammatory disease of the middle ear (otitis media) and congenital anomalies such as non-syndromic deafness due to genetic factors [
1] as the most common causes. Age-related hearing loss is most commonly bilaterally symmetrical, progressive, and irreversible. No preventative therapy exists and symptomatic treatment usually involves the use of hearing aids or surgical implantation of cochlear implants.
The molecular mechanisms that underlie the development of hearing loss and individual variation in risk are poorly elucidated. Twin studies have demonstrated a moderate-to-high heritability for age-related hearing loss of 40–70% [
2‐
6] and several genetic factors and genes have been linked to rare non-syndromic congenital hearing loss. However, the genetic factors that influence variation in risk, emergence, and progression of age-related hearing loss in the general population are poorly characterized.
Over the last decade, genome-wide association studies (GWAS) have been able to illuminate the genetic architectures behind a number of human traits and diseases. Recently, genetic data have become available from large cohorts with standardized collection of many phenotypic traits. For example, in the UK Biobank, a cross-sectional cohort of more than half a million United Kingdom (UK) residents, information on hearing-related traits was collected by touchscreen questionnaire as well as speech-in-noise hearing tests. A recent GWAS of self-reported hearing difficulty and hearing aid use in ~ 250,000 UK Biobank participants identified 44 genetic loci that were associated with these traits and were able to demonstrate the utility of self-reported data for identification of hearing loss-associated genetic variants. In addition, immunological staining revealed murine cochlear expression of three novel hearing-associated proteins: nidogen-2 (encoded by
NID2), clarin-2 (
CLRN2), and the Rho guanine nucleotide exchange factor 28 (
ARHGEF28) [
7].
The cochlea is the critical organ for the sense of hearing and several structural and metabolic components of the human cochlea have been linked to rare forms of syndromic and non-syndromic deafness, such as the gap-junction protein connexin 26 (encoded by
GJB2) [
8] and prestin (
SLC26A5), the motor protein of the outer hair cells [
9]. Imaging of these components in cochlear tissue samples has provided substantial insights into their role in hearing. By necessity, imaging studies have defaulted to laboratory animals due to the paucity of human samples. However, differences in anatomy and molecular function of cochlear components between mammals can complicate extrapolation of results to humans.
In this study, we visualize human hearing-associated proteins, identified as candidates by GWAS, in human cochlear samples by immunohistochemical staining combined with super-resolution structured illumination microscopy (SR-SIM). Human cochlear samples were collected during surgical removal of life-threatening petroclival meningioma and directly fixed for optimal preservation and antigenicity of protein structures.
Discussion
GWAS of self-reported hearing loss-related traits in the UK Biobank revealed that a substantial amount of the associated loci, approximately 40%, could be linked to genes that were determined to be biologically relevant for human hearing. Furthermore, associations overlapped with human congenital non-syndromic deafness loci and associations were found that were proximal to genes with determined functional roles in human auditory mechanotransduction, such as SLC26A5, which encodes the motor protein of the outer hair cells, prestin. A large number of associated loci also contained genes that have previously been linked to hearing-related traits and cochlear structures in vitro and in vivo in other species. Imaging analyses of GWAS-associated genes in human tissue samples are uniquely informative due to the discrepancies in cochlear morphology and protein expression that occur between species. In addition, we were also able to confirm the cochlear expression patterns of several hearing-associated proteins for the first time in human tissue and generate precise mapping of protein expression even at the subcellular level.
Taken together, our findings strongly implicate spiral ganglion function to be critical for hearing loss. Nine of the proteins that were selected for staining were found to be expressed in human type I spiral ganglion neurons: UBE3B, MMP2, PTK2/FAK, GRM5, SYNJ2, LMX1A, EYA4, NOX4, and TRIOBP. Additional genes with neuronal function were also identified from GWAS:
DYRK2 (Dual specificity tyrosine-phosphorylation-regulated kinase 2),
RAB27B (Ras-related protein Rab-27B),
SEPT11 (Septin-11),
NTRK3 (NT-3 growth factor receptor),
ASTN2 (Astrotactin-2),
BAIAP3 (BAI1-associated protein 3), and
MAST4 (microtubule-associated serine/threonine-protein kinase 4). Similar to the hair cells, spiral ganglion neurons are sensitive to environmental factors such as noise, ototoxic compounds, and inflammation [
139]. Damage to the spiral ganglion cells can also occur without apparent damage to the hair cells. In mice, a reversible noise-induced shift in the auditory threshold led to acute loss of synaptic terminals at the inner and outer hair cells [
140], which is likely to be caused by glutamatergic excitotoxicity that is induced by hair cell hyperstimulation. Otoacoustic emissions and histological imaging revealed structurally intact hair cells following exposure [
140]. However, a delayed degeneration of spiral ganglion neuronal cells along with progressive hearing loss was observed in exposed mice over the following years [
140]. The analogous situation in humans would constitute noise exposure-induced threshold shifts during adolescence, for instance after a loud concert, without apparent hearing loss or hair cell damage after recovery, followed by progressive hearing loss that debuts in middle age. Further animal studies have suggested neurotrophin-3 to protect against this type of noise-induced synaptic damage [
141], and to facilitate synaptic recovery following noise exposure-induced hearing loss [
142,
143]. In light of these observations, it is interesting to note that we observed variants linked to neurotrophic signaling pathways to be associated with hearing aid use:
NTRK3, which encodes the receptor for the neurotrophic factor neurotrophin-3 (NT-3) and has previously been found to be expressed in the spiral ganglion of rat [
144‐
147] and chicken [
148], as well as
RAB27B (Ras-related protein Rab-27B), which mediates anterograde axonal transport of the BDNF/NT-3 growth factor receptor, TrkB [
149]. Associations were also observed that could be linked to genes involved in biological processes related to neuronal plasticity, such as extracellular remodeling (
MMP2), neuronal migration (
ASTN2 and
FAK), and cytoskeleton organization and neurite outgrowth (
DYRK2,
NOX4, and
SEPT11).
Our findings also suggest that structural resilience of the human hair cells, and perhaps more critically the sterocilia, is an important factor for the risk of human hearing loss that is affected by common genetic variation. Imaging of GWAS-associated genes in human cochlear tissue showed the actin-associated proteins LMO-7, TRIOBP, and NOX4 to all be expressed in the human hair cells, the hair cell cuticulae, and stereocilia. In addition, five additional structural regulators that are expressed in the hair cells, stereocilia, and cuticula in other species were also associated with hearing-related traits in our analyses:
SPTBN1 (beta-II spectrin) [
78],
CDH13 (cadherin-13) [
95],
BAIAP2L2 (brain-specific angiogenesis inhibitor 1-associated protein 2-like protein 2) [
108],
FSCN2 (fascin-2) [
150], and
TPRN (taperin) [
97]. Age-related hearing loss has been hypothesized to involve the accumulated damage to the stereocilia and hair cells due to damaging environmental exposures in an individual’s lifetime [
151]. Unlike other non-mammalian species, e.g., reptiles and birds, humans are not able to regenerate lost hair cells, e.g., through differentiation of supporting cells to form new hair cells [
152]. Survival of the hair cells is thus critical for maintaining hearing across the lifespan. Our results indicate that differences in resilience to structural damage, e.g., from noise, ototoxic compounds, or inflammatory processes, form part of the underlying genetically determined variation in risk for and development of human age-related hearing loss.
Associations linked to SYNJ2 and GRM5 point towards the cochlear conductance of auditory signals to play a role in hearing loss risk. Both genes encode components of neuronal signaling: GRM5 encodes the metabolic glutamatergic receptor 5, an excitatory G protein-coupled receptor that is activated by glutamate in the synaptic cleft, and SYNJ2 encodes synaptojanin 2, which mediates presynaptic recovery of secretory vesicles after neurotransmitter release. Both proteins were also expressed in type I spiral ganglion neurons. SYNJ2 was also expressed in the outer spiral bundles, i.e., the synaptic connections between the auditory neurons and the outer hair cells.
Our findings also suggest that genetic variants that affect cochlear development and maintenance of cellular identity also contribute to the risk for age-related hearing loss. LMX1A and EYA4 are both transcriptional regulators that have been linked to the development and maintenance of cochlear structures. Our imaging reveals expression of both proteins in discrete structures of the adult human cochlea: within the spiral ganglion neuronal cells, EYA4 within cells of the epithelial layer that faces the scala media, and LMX1A in the organ of Corti. This suggests a role for these proteins in maintenance of the cellular state of these structures in the adult human cochlea.
Limitations
The current study is limited by the lack of audiometric data collected in a standardized manner, which would allow for a more optimal assessment of hearing loss in study participants. Collection of such data require specialized equipment such as a sound-proof room, tone generator, and a trained audiologist for accurate measurements, and thus become difficult to perform on the scale that is commonly required for an adequately powered GWAS for complex traits. Self-report data are valuable proxies for hearing loss as the collection of these data is easily scaled. The large number of biologically relevant findings from GWAS, as determined from literature analyses and immunohistochemical staining of human cochlea, also serve as a validation of the utility of self-report data on hearing-related traits for large-scale studies. However, the type of hearing loss that UK Biobank participants suffer from cannot be determined from questionnaire data alone. Although age-related hearing loss can be considered the most common form of sensorineural hearing loss in the elderly (> 60 years) [
153], other types of hearing loss are common at middle age, such as noise-induced hearing loss and hearing loss due to viral or bacterial infections or ototoxic medications, as well as otosclerosis. The presence of other types of hearing loss may lead to decreased statistical power due to etiological heterogeneities in the studied phenotypes. The results from the GWAS would therefore benefit from replication in additional cohorts with more well-characterized patients.
We are also unable to elucidate the genomic mechanisms that underlie the associations observed in the GWAS. We have instead used a bioinformatic approach to generate plausible mechanisms for associated loci and to select candidate genes for downstream experiments. For hearing-related traits, many associations were found at genetic loci that contain genes with established links to hearing, e.g., at human autosomal nonsyndromic deafness loci, or near genes that affect cochlear morphology in other species. We consider this type of overlap as strong evidence for these genes to be involved in the causal mechanisms that underlie the association with hearing loss. In vitro and in vivo modeling of the effects of genetic variants would allow for determining the causal variant or variants and provide more exact genomic models for how they affect hearing loss risk.
Our study also does not fully account for genetic effects that may differ between sexes or are specifically expressed only in females or males. A recent study on genetic data from the UK Biobank demonstrated that sex-specific effects in the genetic architecture can lead to issues when analyses are performed on a combined set of male and female participants, such as genetic effects being masked from detection due to differences in effect sizes or directions between sexes [
154]. This was most prominent for genetic variants that are associated with distribution of body fat and certain inflammatory disease such as gout and ankylosing spondylitis and the study did not report differences in genetic effects for clinically diagnosed hearing loss (data field 41270: ICD10-diagnoses - H91 Other hearing loss (
n = 10,175) and ICD10-H90 Conductive and sensorineural hearing loss (
n = 2927)). Sex-specific effects may nonetheless be present for hearing loss-related traits, which can potentially be identified in stratified analyses.
Conclusion
In conclusion, GWAS for hearing loss-related traits reveal a genetic architecture for hearing loss in humans. Imaging of GWAS-identified leads findings provide evidence that implicate processes of the spiral ganglion neuronal cells, such as neuronal plasticity and recovery after trauma, to be critical for hearing loss risk. The results also highlight the importance of structural components of the hair cells and hair cell stereocilia and suggest that structural resilience within the organ of Corti is an important factor in determining the risk for sensorineural hearing loss. Importantly, proteins were selected for imaging based on previous cochlear imaging studies in other species. Consequently, it is essential to consider other aspects related to cochlear function that may also be influenced by common hearing loss-associated genetic variation. These findings provide biological insight of the underlying mechanisms of age-related hearing loss among the general population and also offer interesting leads for the development of potential novel therapeutic targets.
Methods
Study cohort
This research has been conducted using the UK Biobank resource under application number 26808. The UK Biobank is a large cross-sectional cohort of almost half a million participants from the UK that were recruited between 2006 and 2010 [
155]. Participants were assessed at 22 centers across the UK where they provided biological samples and answers to an extensive touchscreen questionnaire and underwent physical examinations. At assessment, participants were between aged 37 and 73 years. The cohort was filtered for participants who self-identified as Caucasian and British and were identified as Caucasian by genetic principal component analysis, to avoid eventual issues caused by population stratification. Participants were also filtered to remove participants that exhibited high genetic relatedness (i.e., closely related), poor genotype call rate (< 95%), and high heterogeneity. Participants with discrepancies between genetic and self-reported sex, or a high rate of missing genetic markers were also removed. After filtering, 362,396 participants remained for analyses.
Phenotypes
Hearing aid use, hearing difficulty, and tinnitus were assessed via touch-screen questionnaire. Participants were asked “Do you use a hearing aid most of the time?” and could answer “Yes” or “No” (data field 3393). For hearing difficulty, participants were asked “Do you have any difficulty with your hearing?” and could respond “Yes,” “No,” or “I am completely deaf” (data field 2247). For tinnitus, participants were asked “Do you get or have you had noises (such as ringing or buzzing) in your head or in one or both ears that lasts for more than 5 min at a time?,” and could answer “Yes, now most or all of the time,” “Yes, now a lot of the time,” “Yes, now some of the time,” “Yes, but not now, but have in the past,” or “No, never” (data field 4803). Hearing aid use and hearing difficulty responses were coded as 1 (“Yes”) or 0 (“No”). Participants who declined to answer or answered “I am completely deaf” were set as missing. For tinnitus, “Yes, now most or all of the time,” “Yes, now a lot of the time,” “Yes, now some of the time,” or “Yes, but not now, but have in the past” were coded as 1 and “No, never” was set as 0.
Speech-in-noise perception was assessed during the touchscreen questionnaire session. The participant was supplied headphones and assessed for the ability to hear three spoken digits alongside a rushing background noise [
156]. Participants’ left and right ears were tested separately. For each ear, fifteen triplets of numerical digits along with background noise are presented to the participant. The signal-to-noise ratio ranges from − 12 dB to + 8 dB. The test aims to establish the speech reception threshold, i.e., the signal-to-noise ratio at which the participant is able to comprehend half of the spoken information. It accomplishes this by raising the noise level after a correct response, and lowering it after an incorrect one. The speech recognition threshold for each ear correspond to the value of the signal-to-noise ratio at the last round of the test for participants who completed all 15 rounds [
157] and is provided for the left and right ear (data fields 20019 and 20021). Speech-in-noise was recoded as the mean speech recognition threshold for both ears. Mean thresholds were rank transformed using the “rntransform” function, which is included in the GenABEL package [
158], in R [
159] while adjusting for age and age-squared.
Genotyping
UK Biobank participants were genotyped on two separate microarrays. Initially, approximately 50,000 participants were genotyped on the Affymetrix UK BiLEVE Axiom array as part of the UK BiLEVE (UK Biobank Lung Exome Variant Evaluation) study [
160]. Approximately 450,000 participants were subsequently genotyped on the Affymetrix UK Biobank Axiom array. The two microarrays each contain probes for 807,411 and 825,927 markers, respectively, with 95% overlap of marker content between arrays [
161]. Imputation predicts the genotypes that were not directly assayed in a population. This is achieved by utilizing densely genotyped data sets as reference. In the UK Biobank, imputation of up to 92,693,895 SNPs, insertion-deletions, and large structural variants was carried out using a merged reference set from UK10K, 1000 genomes, and the haplotype reference consortium [
161,
162].
Genome-wide association study, GWAS
Genome-wide association tests were performed with Plink v1.90b3n [
116]. Hearing aid use, hearing difficulty, and tinnitus were analyzed by logistic regression modeling, and speech-in-noise was analyzed by linear regression modeling. Models included sex and age as covariates, as well as a batch variable to adjust for the genotyping platform (Affymetrix UK BiLEVE Axiom array or Affymetrix UK Biobank Axiom array). Models also included the first 15 genetic principal components that were provided by UK Biobank to adjust for any residual effects of population stratification. Genotypes were filtered for minor allele frequency (MAF > 0.01%), genotype call rate (> 95%), and deviation from Hardy-Weinberg equilibrium (
P < 1 × 10
−20). This left approximately 35 million SNPs for association testing. Association analyses were performed on the computational cluster at the Uppsala multidisciplinary center for advanced computational science (UPPMAX) under project sens2017538.
Clumping
Genome-wide association analyses results in a large number of associations between genetic variants and the interrogated phenotype. Associations are commonly seen for many SNPs in close proximity of each other, due to the linkage disequilibrium between SNPs. To determine the number of independent associations, we perform “clumping,” where associated SNPs at each locus are clumped together, and an independent “lead SNP” is identified. Clumping is included as a function in Plink v1.90b3n under the --clump flag and was performed for all traits. We set a lenient threshold for genome-wide significance in the clumping analyses: “--clump-p1 5 × 10−7” for the lead SNP. We used a lenient threshold for correlation between SNPs, “--clump-r2 0.1”, and “--clump-kb 500,” so that all SNPs within 500 kb with a squared correlation of R2 > 0.1 to the lead SNP were assigned to the clump that was represented by the lead SNP. A secondary significance threshold for clumped SNPs of “--clump-p2 5 × 10−7” was also used to designate genome-wide significant SNPs within the clump. We also used a hg19 gene range list to cross-reference and report the overlap between clumps and genes, which is performed with the “--clump-range” flag.
Linkage disequilibrium (LD) score regression
LD score regression (LDSC, v.1.0.0) [
163] was used to generate heritability estimates, genomic inflation factors
(λGC), mean
χ2 statistics, and LD-score regression intercepts for each GWAS. LD scores from a random subset of 5000 UK Biobank participants were used as weights to adjust for correlation between SNPs. LD scores were generated from genotype data in the Plink -bed format with the LDSC software using a 1000-kb window around each SNP. Mean
χ2 represents the mean of the
χ2 statistics for all variants that were tested for association in the GWAS. One minus the intercept of the
χ2 statistics regressed against the LD score (LD score regression) provides an estimate of the mean contribution of the confounding bias in the test statistic. This can be presented as a ratio or the proportion of the
χ2 statistic that the LD score regression intercept ascribes to causes other than polygenic heritability (Ratio = LD intercept − 1)/(mean(
χ2) − 1). In practice, the ratio is commonly around 10–20% [
163]. The GWAS summary statistics can be adjusted for inflation by dividing the
t-statistic with the genomic inflation factor and recalculating the
p values.
Cross reference of GWAS data with functional annotations from databases
To determine possible causal variants and link the associated loci in our GWAS to candidate genes, we cross-referenced our GWAS results with functional annotations from publicly available databases. Databases used were dbSNP build 150 (
www.ncbi.nlm.nih.gov/snp) [
164], the genotype tissue expression database (GTEx,
gtexportal.org) [
165], and the GWAS Catalog (
https://www.ebi.ac.uk/gwas, dataset accessed 1 Mar 2019) [
166]. dbSNP was filtered for non-synonymous missense, nonsense, and frameshift variants before cross-reference. Recent GWAS that had yet to be included in the GWAS Catalog were cross-referenced manually.
The linkage disequilibrium (LD) pattern for each lead SNP from our GWAS was determined using the “--r2” command with the flags “--ld-snp,” “--ld-window-r2 0.1,” and “--ld-window-kb 2000” in Plink v1.90b3n. All SNPs within a 2-Mb window around each lead SNP that were in linkage disequilibrium with the lead SNP (
R2 > 0.1) were then identified and annotations for all SNPs were extracted from dbSNP and the GWAS catalog. SIFT and PolyPhen2 scores for nonsense, missense, and frameshift mutations were identified manually from Ensembl (
www.ensembl.org). SIFT and PolyPhen are tools for predicting the functional consequences of genetic variants within protein-coding regions based on sequence homology as well as the physicochemical properties between the alternate amino acids [
109,
110].
For the GTEx database, we determined the correlation of the lead SNPs from our analyses with expression quantitative trait loci (eQTLs) that were included in GTEx. We used a significance threshold for association of genetic variants with gene expression in GTEx of
P < 2.3 × 10
−9, in line with previous studies [
167]. The correlations between the lead SNPs from our GWAS analyses and eQTLs from GTEx were determined and genetic variants in LD (
R2 > 0.8) were considered to overlap.
Literature review
We wished to determine if any genes within our GWAS-identified loci had been previously linked to hearing or hearing-related traits in humans or other species. We therefore performed a PubMed search of all genes within the clumps that were determined by Plink combined with the terms “AND Hearing.” The search results were then manually filtered for original peer-reviewed research articles that had reported data on at least one of the genes included in the literature search.
Collecting and processing human tissue
The surgical specimens were from patients suffering from life-threatening posterior cranial fossa meningioma compressing the brain stem. The operations were performed at Uppsala University Hospital by a team of neurosurgeons and oto-neurosurgeons. Human cochleae were dissected out using diamond drills of various sizes. Tissues were immediately placed in 4% paraformaldehyde diluted with 0.1 M phosphate-buffered saline (PBS; pH 7.4) in the operating room. After a 24-h period spent in fixative, specimens were washed in 0.1 M PBS and then placed in 10% Na-EDTA solution at pH 7.2 for decalcification. The Na-EDTA solution was changed every 2 days until the decalcification process was completed, which took approximately 3 to 6 weeks. Decalcified cochleae were rinsed with 0.1 M PBS and placed in 25% sucrose in 0.1 M PBS overnight (4 °C). Cochleae were embedded in Tissue-Tek (OCT Polysciences) for frozen sections. Cochleae were rapidly frozen and sectioned at 8–10 μm using a Leica cryostat microtome. Frozen sections were collected onto gelatin/chrome-alum-coated slides and stored in a freezer of − 70 °C before immunohistochemistry procedures [
168,
169]. In total, five specimens were available for staining. Patient characteristics are presented in Additional file
1: Table 7.
Antibodies and immunostaining procedures
Immunohistochemistry procedures on human cochlear sections were described in previous publications [
170,
171]. Sections were incubated with an antibody solution under a humidified atmosphere at 4 °C for 20 h. Sections were incubated with secondary antibodies conjugated to Alexa Fluor (Thermo Fisher Scientific, Uppsala) counterstained with the nuclear stain 4′,6-diamidino-2-phenylindole dihydro-chloride (DAPI), mounted with ProLong® Gold Antifade Mountant (Thermo Fisher Scientific, Uppsala, Catalog number: P10144), and then covered with the specified cover glass compatible with both confocal and super-resolution microscopes. Primary and secondary antibody controls and labeling controls were used to exclude endogenous labeling or reaction products [
172]. The antibodies used for immunohistochemistry are shown in Additional file
1: Table 8.
Imaging and photography
Stained sections were first investigated with an inverted fluorescence microscope (TE2000; Nikon, Tokyo, Japan) equipped with a spot digital camera with three filters (for emission spectra maxima at 358, 461, and 555 nm). Image-processing software (NIS Element BR-3.2; Nikon), including image merging and a fluorescence intensity analyzer, was installed on a computer system connected to the microscope. For laser confocal microscopy, we used the same microscope equipped with a three-channel laser emission system. SR-SIM [
173] was performed using a Zeiss Elyra S.1 SIM system and a × 63/1.4 oil Plan-Apochromat objective (Zeiss, Oberkochen, Germany), sCMOS camera (PCO Edge), and ZEN 2012 software (Carl Zeiss Microscope). The resolution of the SR-SIM system at BioVis, Uppsala University, was 107 nm in the
X–
Y plane and 394 nm in the
Z plane. The laser and filter were set up as follows: 405 nm laser of excitation coupled with BP 420–480 + LP 750 filter, 488 nm laser of excitation with BP 495–550 + LP750 filter, 561 nm laser of excitation with BP 570–620 + LP 750 filter, and 647 nm laser of excitation with LP 655 filter. From the SR-SIM dataset, 3D reconstruction was done with Imaris 8.2 (Bitplane, Zürich, Switzerland). A bright-field channel was able to merge fluorescence channels to visualize cell and tissue borders.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.