Genome-wide analysis implicates inner ear development in Ménière’s disease

Abstract

Ménière’s disease (MD) is a chronic inner ear disorder characterized by recurrent vertigo, fluctuating sensorineural hearing loss, and tinnitus. Despite these distinctive symptoms, its etiology remains poorly understood. We performed a genome-wide meta-analysis of 8,969 cases and 1,962,542 controls across five large biobanks, identifying five independent genome-wide significant loci and estimating an observed-scale SNP heritability of 7% (SE 0.8%), consistent with a modest but significant genetic contribution to MD risk. Fine-mapping and integrative functional analyses implicate two convergent biological processes – developmental regulation of the inner ear, involving EYA4, EYA1, and LMO4 – and retinoic acid metabolism, with loci near CYP26A1/C1 and ALDH1A2 suggesting disrupted RA signaling in sensory and fluid-pressure homeostasis. These developmental regulator genes are robustly expressed in fetal and adult human inner ear cell types, supporting a model in which altered developmental programs predispose to adult vestibular and auditory dysfunction. Phenome-wide and genetic correlation analyses further reveal shared genetic architecture between MD and related traits, including vertigo, tinnitus, hearing loss, migraine, and sleep apnea, situating MD within a broader spectrum of sensory and neurological disorders. Collectively, these findings establish a genetic framework for Ménière’s disease risk and implicate developmental regulators and retinoic acid signaling as key contributing pathways.

Introduction

Ménière’s disease (MD) is a chronic inner ear disorder characterized by debilitating episodes of vertigo, fluctuating sensorineural hearing loss (SNHL), tinnitus, and aural fullness^1–3. The etiology of MD remains unclear, but it is frequently associated with endolymphatic hydrops, an abnormal accumulation of endolymph within the inner ear⁴. Several pathophysiological mechanisms have been proposed, including fluid imbalance, infection, and vascular dysfunction, but these hypotheses lack direct genetic or molecular support⁵.

MD affects approximately 1 in 2000 individuals, typically presenting in middle to late adulthood, with higher prevalence among women and those of European ancestry^6–10. Genetic studies indicate a heritable contribution to MD risk, with familial cases (5–15% of patients) following autosomal dominant inheritance with ~60% penetrance¹¹. Rare coding variants in genes including OTOG, GJD3, FAM136A, and DTNA have been identified in familial MD^12–21. However, the genetic basis of sporadic MD, which constitutes the majority of cases, remains poorly understood⁵. The contribution of common variants and the proportion of risk they explain have not been systematically defined, leaving the full genetic architecture and downstream molecular mechanisms unresolved.

Recently, large-scale initiatives to expand biobank sample sizes, harmonize phenotypic data, and perform high-quality genome-wide genotyping across diverse populations have transformed the study of complex traits at scale^22–28. These resources enable well-powered genome-wide association studies (GWAS) and meta-analyses, even for low-prevalence disorders^29–31. Furthermore, meta-analyses across biobanks allow for the detection of novel risk loci, refinement of heritability estimates, and mapping of genetic variants to molecular function, thereby providing opportunities to address critical gaps in our understanding of the genetic architecture of MD.

We leveraged genome-wide summary statistics from five major biobanks – UK Biobank (UKB), All of Us (AoU), Million Veteran Program (MVP), FinnGen, and Biobank Japan (BBJ) – to characterize the genetic architecture of MD. Through meta-analysis of 8,969 MD cases and 1,962,542 controls from three populations across these biobanks, and downstream fine-mapping and integrative annotation, we identified five independent genome-wide significant signals at three loci, and a SNP-based heritability of 7%. Four of these signals map to potential regulatory elements for EYA4 and EYA1, genes essential for inner ear development. These associations, combined with suggestive evidence at LMO4, highlight a developmental regulatory mechanism. Additional signals near CYP26A1/C1 and ALDH1A2 implicate retinoic acid signaling, a pathway critical for sensory and fluid-pressure homeostasis. Phenome-wide and genome-wide correlation analyses further demonstrate shared genetic architecture between MD and non-symptom traits, including migraine (r_g=0.28) and sleep apnea (r_g=0.20). Collectively, these findings support a polygenic basis for Ménière’s disease driven by regulatory variation in developmental and morphogen pathways, providing a genetic framework for understanding its complex etiology.

Results

Meta-analysis identified five independent genome-wide significant variants

We analyzed data across five biobanks – All of Us^24,32 (n = 383,735; 841 MD cases), Million Veteran Program²⁷ (n = 510,058; 2,306 cases), UK Biobank²⁵ (n = 420,473; 1,220 cases), FinnGen²⁶ (n = 478,519; 3,680 cases), and Biobank Japan³³ (n = 178,726; 1,290 cases) (Table 1). In aggregate, these datasets include 8,969 MD cases and 1,962,542 controls, corresponding to an overall prevalence of 0.46% in our study population. MD status was defined using phecode 386.1 or an equivalent phenotype mapped from ICD codes (H810, 3860, and 38599), with additional quality control steps applied as described in the Methods. To assess the contribution of common variants to MD genetic architecture and disease risk, we restricted analyses to variants with minor allele frequency (MAF) > 1%, yielding 13,423,260 variants. MD cases were more frequent in females (Figure 1a), and prevalence increased with age (Figure 1c, 1d), consistent with prior reports⁵. Across biobanks, 81.4% of MD cases were of European ancestry (Figure 1b), and cases were enriched for European ancestry within All of Us (Supplementary Figure 1).

Table 1.

The five GWAS performed on independent datasets, with corresponding country of collection and ancestry filtering of cases and controls, number of cases and controls in each study, and number of SNPs after filtering for minor allele frequency greater than 1%.

GWAS	Country; Ancestry	Cases/Controls	Case definition	Data Type	Number of SNPs
All of Us (AoU)	USA; Multi-ancestry	841/382,894	Phecode 386.1	Whole Genome Sequenced	6,795,311
Million Veteran Program (MVP)	USA; European and Admixed American	2,306/507,752	Phecode 386.1	Genotyped and Imputed	9,254,829
UK Biobank (UKB)	UK; European	1,220/419,253	UKB phenocode 20002, coding 1421	Genotyped and Imputed	9,549,936
FinnGen	Finland; European	3,312/474,839	H81.0 (ICD-10); 3869 (ICD-9); 38599 (ICD-8)	Genotyped and Imputed	9,462,981
Biobank Japan (BBJ)	Japan; East Asian	1,290/177,436	Phecode 386.1	Genotyped and Imputed	7,440,824

Figure 1. Demographics of Ménière’s Disease Cases.

a. Sex distribution of Ménière’s Disease cases across FinnGen, All of Us, and UK Biobank. Information on the Million Veteran Program and Biobank Japan was not available. b. Genetically inferred ancestry distribution of MD cases across five contributing studies (Methods). Other genetically inferred ancestries include African, South Asian, and Middle Eastern. c. Age distribution of Ménière’s disease cases in All of Us and FinnGen, categorized by age bins, in proportion. Information on the Million Veteran Program, the UK Biobank, and Biobank Japan was unavailable. d. Age-specific prevalence of Ménière’s disease cases in All of Us and FinnGen.

We conducted a GWAS in All of Us using Regenie, and then performed a fixed-effects meta-analysis with summary statistics from the four other studies^24–27,33 using the inverse-variance weighted scheme in METAL³⁴. For each study, we filtered out variants with minor allele frequency <0.01 or minor allele count <10. We estimated SNP-based heritability to be 7% (SE=0.8%) on the observed scale using LD Score Regression³⁵ (LDSC) applied to the meta-analyzed summary statistics. Minimal genomic inflation (λ_gc = 1.01) in the meta-analyzed result indicated negligible confounding before correction. We identified five independent genome-wide significant index SNPs (P<5×10⁻⁸; Figure 2; Table 2), defined as the most significant variant within a 1 Mb window, clumped at r² > 0.05, and supported by at least one additional genome-wide significant SNP (Supplementary Table 1). Three other loci were just below the significance threshold (Methods; 5×10⁻⁸ < P < 5×10⁻⁷), and we highlight signals near LMO4 and ALDH1A2 due to their biological relevance (Supplementary Table 2, 3).

Figure 2. Manhattan plot of meta-analysis.

GC-corrected −log10 of P values from the meta-analysis of multi-ancestry Ménière’s Disease GWASs. The red dashed line indicates the genome-wide significance threshold of 5e-8. Genome-wide significant variants, where no other significant variant is within 1 Mb, are plotted in grey. The nearest genes to the index SNPs are annotated in black. Additional loci discussed in the text are annotated in grey. Inset: quantile-quantile plot of the meta-analysis with genomic control applied.

Table 2.

The five genome-wide significant index SNP associations, with corresponding RsID, location in hg38, effect allele, frequency of effect allele in 1000 Genomes EUR population, odds ratio, P-value, and closest gene by distance.

RsID	Location (chr:position)	Effect Allele (EA)	Frequency of EA in 1000 Genomes EUR	Odds Ratio	P-value	Closest Gene
rs11752136	6:133184618	A	0.23	0.852	2.71×10−14	EYA4
rs3777811	6:133394072	G	0.09	1.246	1.34×10−17	EYA4
rs2639899	8:71007227	C	0.47	0.905	7.25×10−9	EYA1
rs145240113	8:71583477	G	0.10	0.852	1.08×10−9	EYA1
rs61861121	10:93191324	A	0.09	1.197	9.51×10−13	CYP26A1

Two independent signals at EYA4 suggest a developmental contribution

We next examined the two independent association signals located 209 kb apart at EYA4 (index SNPs rs11752136 and rs3777811, r² = 0.003; Table 2; Figure 3a). EYA4 encodes a member of the evolutionarily conserved eyes absent (EYA) protein family and functions as both a transcriptional coactivator and a phosphatase. EYA4 has essential roles in auditory development and function and loss-of-function mutations lead to autosomal dominant sensorineural hearing loss (DFNA10)^36–40. Significant SNPs clumped with rs11752136 overlap candidate cis-regulatory elements located ~60kb upstream of EYA4 (Encode track, UCSC browser), the EYA4 transcription start site (TSS), the first intron, and an enhancer in intron 1. rs3777811 is located 155 kb downstream of the EYA4 TSS in intron 3 (Figure 3a). Fine-mapping with SuSiE⁴¹ resolved the region into two credible sets corresponding to the two index SNPs; however, high LD precluded confident identification of a single causal variant in either region (Figure 3b; Supplementary Table 4–6). EYA4 shows a strong gene-level association with MD (MAGMA P=3.3×10⁻¹⁰; Supplementary Table 10).

Figure 3. Characterization of genome-wide significant loci for Ménière’s disease.

(a) −log₁₀ P-values from the meta-analysis plotted against genomic positions for each index SNP region. The window size of the plots is 500kb. Genes of interest are colored in red along the x-axis. Red dotted line denotes genome-wide significance (P<5×10⁻⁸). (b) Posterior inclusion probabilities (PIPs) from SuSiE fine-mapping plotted against genomic position for each index SNP region. Shading denotes the start and end of credible sets, and the number of SNPs in the credible set is annotated. (c) Forest plots of odds ratios as OR for each GWAS study and meta-analysis with 95% confidence intervals on the x-axis for each index SNP. For each index SNP, effective allele and heterogeneity P-values are included in the title.

EYA4 is expressed in sensory and non-sensory cell types of the cochlea and vestibular system at fetal and adult stages according to single-cell RNA-seq data from the human inner ear (Figure 4). Although eQTL data are not currently available for the inner ear, rs3777811 is an eQTL for EYA4 in several brain tissues, with the MD risk allele associated with increased EYA4 expression (Supplementary Table 11), whereas weak LD proxies of rs11752136 (rs971589; r² = 0.48) are eQTLs for EYA4 across brain, heart, and tibial nerve tissues (Supplementary Table 11). Phenome-wide analysis showed that rs3777811 is associated with increased risk of sensorineural hearing loss (SNHL), tinnitus, and general ear disorders, core symptoms that are genetically correlated with MD (Figure 5a, b), as well as decreased risk of sleep apnea (Supplementary Table 16). Notably, rs3777811 is distinct from, and not in LD with, the previously identified coding variant rs9493627, which is associated with SNHL⁴². In contrast, rs11752136 is an MD-specific signal and shows no association with related phenotypes (Supplementary Table 15). Together, these complementary signals suggest distinct regulatory mechanisms, one shared with SNHL and one specific to MD. Functional characterization of these two variants may help identify molecular mechanisms that are both specific to MD and shared with related disorders.

Figure 4. Dotplot of single-cell expression from human inner ear scRNAseq across developmental stages.

Dot plot shows the expression of EYA4, EYA1, CYP26A1, CYP26C1, and LMO4 across diverse human inner ear cell types at three developmental stages (fetal week 7.5, fetal week 9.2, and adult). Dot size represents the percentage of expressing cells, and color indicates scaled average expression. Data were obtained from the human inner ear atlas⁶⁰ via the gEAR Dashboard⁶¹.

Figure 5. Phenotypic characterization.

(a) Phenome-wide associations from harmonized FinnGen, MVP, and UKB GWAS results of the five index SNPs across disease categories are plotted. The red dotted line denotes the Bonferroni-corrected P-value threshold. The top plot indicates the same direction of effect as Ménière’s disease, while the opposite direction is indicated in the bottom plot. Phenotypes with significant association are denoted. (b) Forest plot of genetic correlation estimates of Ménière’s disease between phenotypes across categories. Asterisk denotes Bonferroni-corrected significance.

Two independent signals at EYA1 further emphasize the role of development

At EYA1, we found another two independent index SNPs, rs2639899 and rs145240113 (r²= 1×10⁻⁴; Table 2; Figure 3a). rs2639899 lies 192 kb downstream of EYA1 in a candidate cis-regulatory element within a genomic region on chromosome 8 that also features a cluster of five ultraconserved non-coding elements (UCNEBase track, UCSC Genome Browser). rs145240113 is located 35 kb upstream of the EYA1 TSS. Fine-mapping delineated two credible sets corresponding to these index SNPs, though again, extensive local LD precluded identification of causal variants (Figure 3a, 3b; Supplementary Table 4, 7, 8).

EYA1’s association with MD reached gene-level significance (MAGMA P=1.3×10⁻⁵; Supplementary Table 10), and similar to EYA4, is expressed in the human inner ear at fetal and adult stages (Figure 4). EYA1 encodes a homologous dual-functioning transcriptional regulator and phosphatase, and its haploinsufficiency causes Branchio-Oto-Renal (BOR) syndrome, a congenital disorder characterized by malformations of the outer, middle and inner ear, hearing loss, and renal anomalies^43–45. Although EYA1 and EYA4 share similar phosphatase substrates³⁶, EYA1 functions earlier in otic development than EYA4, with EYA1-deficient ears arresting at the otic vesicle stage⁴⁶.

rs2639899 is an eQTL for EYA1 in esophageal mucosa. Consistent with the effect of rs3777811 on EYA4, the MD risk allele of rs2639899 is associated with increased EYA1 expression (Supplementary Table 12). Besides Ménière’s disease, the risk-increasing allele of rs2639899 is associated with several phenotypes, including reduced risk of glaucoma, while the risk-increasing allele of rs145240113 is associated with increased risk of hyperparathyroidism and decreased risk of SNHL (Figure 5a; Supplementary Table 17, 18).

An association at LMO4 is consistent with its developmental role in mice

We observed a suggestive association at rs1199565, located 101 kb downstream of LMO4 (P = 1.46×10⁻⁷; Supplementary Table 2; Supplementary Figure 2) and annotated as an eQTL in subcutaneous adipose tissue (Supplementary Table 13). Although below genome-wide significance, this signal is noteworthy given the established role of LMO4 in vestibular and cochlear morphogenesis: loss of LMO4 expression in mice disrupts inner-ear development and impairs balance^47,48. In the human fetal inner ear, single-cell RNA-seq data show strong LMO4 expression in vestibular epithelial and cochlear duct floor cells (Figure 4), suggesting a conserved developmental function. Together with the EYA4 and EYA1 loci, this locus reinforces a broader theme in which dysregulation of genes guiding inner-ear formation contributes to Ménière’s disease susceptibility.

The CYP26A1/C1 locus implicates retinoic acid signaling

On chromosome 10, the index SNP rs61861121 lies downstream of CYP26A1, CYP26C1, EXOC6, and MYOF (Table 2; Figure 3a, 3b). Fine-mapping identified rs61861121 as a causal SNP within a two-variant 95% credible set, with a posterior inclusion probability of 0.90 (Figure 3b; Supplementary Table 4, 9). rs61861121 is not an eQTL for any nearby genes, and none reached gene-level significance in MAGMA (CYP26A1 P = 0.03; CYP26C1 P = 0.05, EXOC6 P=0.12, MYOF P=0.05; Supplementary Table 10), leaving the causal gene unresolved. Nonetheless, functional considerations suggest that CYP26A1/C1 are plausible candidates. Both genes encode enzymes that degrade retinoic acid to maintain spatial gradients during embryogenesis, a process essential for patterning of the retina, hindbrain, and otic placode^49–52. The functional paralog of CYP26A1/C1 in mice – CYP26B1 – has a critical role in vestibular sensory epithelia formation by shaping retinoic acid availability in the embryonic inner ear⁵³. The vestibular structures affected are responsible for detecting head motion, directly relevant to MD. Although rs61861121 shows no association with other phenotypes after Bonferroni correction (Figure 5a), a more lenient threshold (Methods) suggested associations with multiple eye and ear phenotypes, including glaucoma, peripheral retinal degeneration, and sensorineural hearing loss (Supplementary Table 19). This convergence suggests a role for retinoic acid signaling in a broader category of fluid tension disorders spanning glaucoma, intracranial hypertension, and MD⁵⁴. The potential role for retinoic acid signaling is further supported by a suggestive association at ALDH1A2 (rs6493979; P = 8.6×10⁻⁸; Supplementary Table 2; Supplementary Figure 2), which is MAGMA significant (P = 3.07×10⁻⁷; Supplementary Table 10) and has moderate expression in vestibular epithelial cells (Supplementary Figure 3). ALDH1A2 is directly involved in retinoic acid synthesis^55–58 and associated with cerebral ventricle volume⁵⁹, which, again, points to disruption of retinoic acid metabolism in MD etiology and a link with other fluid tension disorders.

Discussion

In this study, we report the first large-scale genome-wide association study of Ménière’s disease (MD), integrating data from nearly two million individuals across five major biobanks. An observed-scale SNP heritability of 0.07 indicates a measurable but modest contribution of common variants to disease risk. Common variation in genes associated with familial MD does not appear to make a substantial contribution to sporadic cases. We identify five genome-wide significant loci with functional and phenotypic evidence implicating developmental regulation of the inner ear, and retinoic acid–mediated signaling. These findings expand our understanding of the genetic basis of MD beyond familial rare variants and point to biological pathways important in disease.

The most robust associations map to EYA4 and EYA1, transcriptional regulators essential for otic placode and inner ear development^{38–40,44–46}. Both loci harbor two independent non-coding signals. Two of the four index SNPs – rs11752136 at EYA4 and rs145240113 at EYA1 – act as eQTLs in various tissues, and MD risk alleles of both are associated with increased expression of EYA4 and EYA1 (Supplementary Table 11, 12). Meanwhile, a significant association, rs3836957, located at EYA4 and clumped with rs11752136, was identified as an enhancer (Supplementary Table 20). These patterns suggest that MD-associated alleles modulate gene expression rather than causing loss of function. Expression analysis of human inner ear single-cell RNA-seq data further supports the relevance of these genes: both EYA4 and EYA1 are expressed in vestibular and cochlear epithelial cell types during fetal and adult stages, while LMO4 is expressed in vestibular and cochlear duct floor cells (Figure 4). These findings converge on developmental dysregulation of sensory epithelia as a shared mechanism linking MD to congenital and acquired hearing disorders. Consistent with this model, EYA4 variants have been implicated in autosomal dominant sensorineural hearing loss^36,37,40, EYA1 is causal in Branchio–Oto–Renal syndrome⁴⁵, and loss of LMO4 expression leads to maldevelopment of the mouse inner ear^43,44,47,48. The extension of these developmental pathways to MD, typically a late-onset, sporadic condition^2,5, suggests that subtle regulatory perturbations in the expression of developmental genes can predispose to adult vestibular and auditory dysfunction, in agreement with a recently proposed model of MD pathophysiology⁶².

A complementary biological theme arises from the association near CYP26A1/C1 and the suggestive signal at ALDH1A2. These genes encode enzymes that respectively degrade and synthesize retinoic acid (RA), maintaining local RA gradients essential for hindbrain and otic patterning^{50–52,63–66}. Fine-mapping highlights rs61861121 as a likely causal variant with high posterior probability, pointing to regulatory perturbation of CYP26A1/C1 rather than protein-altering effects. In mice, the paralog Cyp26b1 controls vestibular sensory epithelium formation through spatial RA regulation, directly linking this pathway to balance⁵³. These associations align with prior hypotheses connecting RA metabolism to endolymphatic homeostasis and provide a developmental-to-physiological bridge for MD pathogenesis^62,63.

Phenome-wide association analyses and genome-wide genetic correlations highlight the intersection between MD and related sensory or neurological traits. The strongest polygenic overlaps are observed with vertigo, tinnitus, and sensorineural hearing loss, which are core symptoms of MD. We also observe a positive genetic correlation with migraine, a comorbidity of MD that may share vascular or neurogenic mechanisms⁶⁷. However, despite a positive genetic correlation, GWAS signals for vertigo and migraine do not overlap with MD^68,69, indicating that the largest effects are not shared. We also find no evidence that common variants at genes associated with familial MD contribute to risk of sporadic MD. On the other hand, we find overlap in GWAS signals but low genetic correlations for fluid-tension related traits, including glaucoma, suggesting a shared biological pathway, albeit one that only explains a small proportion of risk.

Together, our results establish MD as a polygenic disorder influenced by regulatory variation in genes governing inner ear development and retinoic acid signaling. These factors only explain a small fraction of risk, and it remains to be seen how they interact with other genetic factors or environmental triggers. These findings lay the groundwork for future functional studies using human inner ear organoids and in vivo models to test how subtle dysregulation of pathways involving EYA1/4 and CYP26A1/C1 pathways affects sensory epithelia formation and fluid balance. Beyond mechanistic insight, this work provides a foundation for constructing polygenic risk models and exploring comorbidity across sensory and neurological disorders. Expanding these analyses to larger and more ancestrally diverse cohorts and integrating inner ear–specific regulatory eQTL data will be the next steps toward a comprehensive understanding of MD genetics and pathology and its place within the broader landscape of auditory and vestibular disease.

Methods

Contributing Biobanks and Genome-wide Association Studies (GWAS)

All of Us Research Program

The All of Us Research Program (AoU) is a US-based biobank with genomics data, electronic health record (EHR), and survey information from diverse populations²⁴. We performed a GWAS of Ménière’s disease using Regenie³¹, including age, sex, age2, age2*sex, and the top 10 PCs as covariates. We obtained genotypes from short-read whole-genome sequencing in the Curated Data Repository (CDR) v8 release, and defined phenotypes using phecode 386.1. We filtered variants by population-specific allele frequency > 1% or allele count > 100 in any ancestry subgroup. We performed additional quality control in PLINK 2 using --maf 0.01, --geno 0.02, and --mind 0.05 across all individuals. Related individuals were removed. We defined cases as participants with ≥ 2 phecode encounters, and controls as those with none.

Million Veteran Program

The Million Veteran Program (MVP) is a large and ancestrally diverse U.S. biobank linking genomic data to detailed EHR and survey information⁷⁰. We used summary statistics done by MVP from the GIA gwPheWAS release²⁷, where the Ménière’s disease GWAS (phecode 386.1) was performed with SAIGE⁷¹, among European and Admixed American participants (2,306 cases, 507,752 controls). Cases were defined as participants with ≥ 2 phecode encounters, and controls as those with none. Ancestry was inferred via PCA-based random-forest classifiers. Genotypes were imputed using a hybrid reference panel of African Genome Resources + 1000 Genomes Phase 3 v5 and filtered by imputation quality > 0.3, minor allele count > 20, call rate > 0.975 for MAF > 0.01, and > 0.99 for MAF < 0.01²⁷.

UK Biobank

The UK Biobank (UKB) is a prospective cohort of ~500,000 participants with genotype and EHR data from the United Kingdom²³. We used the Pan-UK Biobank European-ancestry GWAS (phenocode 20002, coding 1421, SAIGE model)²⁵, which included 1,220 cases and 419,253 controls. Genetic ancestry was inferred using PCA-based random-forest classifiers, and only unrelated individuals (PC-Relate) were included. Genotypes were imputed to the UKB version 3 panel (~97 million variants) and filtered by INFO > 0.8 and minor allele count > 20²⁵.

FinnGen

FinnGen is a nationwide research initiative combining Finnish biobank data with health registry records²⁶. We used GWAS summary statistics from the R12 release for endpoint H8_Ménière, analyzed with Regenie. The study included 3,312 cases and 474,839 controls of European ancestry. Phenotypes were harmonized across ICD-10 (H810), ICD-9 (3860), and ICD-8 (38599) codes. Genotypes were imputed to the population-specific SISu v4.2 reference panel and filtered by INFO > 0.9 and MAF > 0.01.

Biobank Japan

Biobank Japan is a large hospital-based biobank that links genomic data with clinical information from ~200,000 Japanese participants across phenotypes^28,72. The BBJ GWAS was conducted using SAIGE to test for associations between ~13 million imputed variants (imputed to the 1000 Genomes Phase 3 East Asian panel) and clinically ascertained diagnoses. For Ménière’s disease, the analysis included 1,290 cases and 177,436 controls, defined by phecode 386.1. Quality control followed the BBJ protocol⁷²: variants with MAF < 0.01, imputation INFO < 0.8, or Hardy–Weinberg P < 1e-6 were excluded, and participants with call rate < 0.98 or excess heterozygosity were removed. Association testing adjusted for age, sex, top 20 genetic PCs, and genotyping batch.

Ancestry Definition

Genetic ancestry assignments for All of Us, the Million Veteran Program, and UK Biobank were obtained from each cohort, where ancestry inference was performed using principal component analysis (PCA) and random-forest classifiers trained on the Human Genome Diversity Project and the 1000 Genomes Project^24,25,27. In FinnGen and Biobank Japan, participants were recruited from genetically homogeneous Finnish and Japanese populations, corresponding broadly to EUR and EAS ancestries, respectively.

GWAS Meta-analysis

We conducted the meta-analysis using METAL with summary statistics from five GWAS described above^{24–27,34,72}. We lifted over Pan-UKB and BBJ summary statistics from GRCh37 to GRCh38 using the UCSC liftOver tool⁷³. We kept single study-private variants, and removed variants with minor allele frequency <0.01 or with minor allele count <10. We applied genomic control to all datasets before meta-analysis and again on the meta-analyzed result. We used the STDERR scheme to perform meta-analysis, utilizing weighted effect size estimates and inverted standard errors. We calculated the effective sample size to replace the overall sample size³⁴. Cochran’s Q and its P-value as measures of heterogeneity were calculated by METAL. We had 14,096,168 SNPs after meta-analysis, and kept a total of 13,423,260 that are also in the 1000 Genomes high-coverage Phase 3 reference panel⁷⁴ for downstream analyses.

Heritability Estimate

We estimated SNP-based heritability using LD Score Regression³⁵ (LDSC). Meta-analysis summary statistics for Ménière’s disease (MD) were first processed using the munge_sumstats.py script provided with LDSC. We excluded strand-ambiguous SNPs and aligned all variants to the 1000 Genome Phase 3 high-coverage reference genome⁷⁴. For LD reference, we computed LD scores using 632 European populations from the 1000 Genomes Project Phase 3 high-coverage data.

We estimated heritability on the observed scale, assuming a case-control design with 8,969 MD cases and 1,962,542 controls. We also attempted, but failed, to convert observed-scale heritability to liability-scale heritability using an assumed population prevalence of 0.2% and the methods of Lee et al. and Ojavee et al.^75,76

Index SNP selection

We defined Genome-wide significant hits at P<5e-8 and suggestive associations at P<5e-7. We obtained index SNPs, or index hits, through LD-clumping using Plink 1.9, with –clump-r2 of 0.05 and –clump-kb of 1 million bps⁷⁷. Genotype data of 632 European individuals from the 1000 Genomes Project Phase 3 high-coverage variants data were used as reference panels⁷⁴. The clumping threshold for P-value is first set to 5e-8 and then loosened to 5e-7 for regions previously having no clumped result. Loci with P values below the thresholds but clumped with no other variants were defined as significant or suggestive loci without support.

Variant Consequences

To predict the consequences of the variants, we employed Ensembl Variant Effect Predictor (VEP)⁷⁸. The input to VEP was the list of SNPs that are significant and clumped with each index SNP.

Fine-Mapping

To localize putative causal variants within genome-wide significant regions, we performed statistical fine-mapping using the Sum of Single Effects (SuSiE)⁴¹. The fine-mapping window was ±500 Kb centered on each index SNP, and the midpoint of two index SNPs on chromosome 6. For each region, SNPs that are present across the five contributing biobanks were kept. We constructed an ancestry-weighted LD panel by effective sample size in the meta-analyzed GWAS from the 1000 Genomes Project Phase 3 high-coverage data⁷⁴. We confirmed accurate LD matching between the reference and meta-analysis data by assessing LD–Z concordance (s ≤ 0.003 across all loci). For each locus, SuSiE was run with up to 10 potential causal components (L = 10) and default prior variance settings. 95% credible sets (CS) were defined as the minimal subset of variants whose cumulative posterior inclusion probabilities (PIP) reached 0.95.

PheWAS of Index SNPs

To assess the phenotypic pleiotropy of lead SNPs from the MD GWAS meta-analysis, we queried publicly available PheWAS results from the FinnGen-MVP-UK Biobank meta-analysis, accessed via FinnGen Data Freeze 12²⁶. For each index SNP, we extracted its association statistics across a curated panel of 330 harmonized phenotypes. Only associations reaching Bonferroni significance of 1.5e-4 were considered statistically significant.

eQTL Characterization

To assess regulatory relevance, we checked whether each index SNP was a significant cis-eQTL in any tissue based on GTEx v10 single-tissue cis-eQTL summary statistics⁷⁹. We considered variant-gene pairs passing permutation-based significance thresholds (FDR < 0.05) across 50 tissues.

MAGMA Gene Association

We performed gene-level association analysis using MAGMA to map SNP-level GWAS meta-analysis summary statistics to gene-level signals, where individual SNP p-values were combined into gene test-statistics⁸⁰. We mapped SNPs in meta-analysis summary statistics to genes through annotation using a protein-coding gene location file in build 38 obtained from the NCBI site, including transcription start site and end site (https://cncr.nl/research/magma/). We adopted an annotation window of 10kb before the transcription start site and 10kb after the end site to include SNPs around the gene. Then, genes were prioritized using the gene annotation file and the meta-analysis summary statistic, where genotype data from the European population in the phase 3 1000 genomes project served as reference data, and a uniform sample size was specified⁷⁴. We identified significant gene associations after Bonferroni correction using the number of annotated genes.

Human Inner Ear Gene Expression

Single-cell RNA-sequencing data were obtained from the human inner ear atlas⁶⁰ and accessed through the gEAR Dashboard (https://umgear.org/). Processed gene expression matrices and cell-type annotations provided by the atlas were used for downstream visualization. All analyses were performed in R using the Seurat package (version 5.3.1). Gene expression values were normalized and scaled using Seurat’s default normalization and scaling workflows. Dot plots were generated using the Seurat function DotPlot.

Genetic Correlation

We estimated genetic correlations using LD Score Regression³⁵ (LDSC). Data processing and LD score calculation were discussed in the section on heritability estimate. To assess shared genetic architecture between MD and other traits, we performed pairwise genetic correlation analysis using the ldsc.py --rg command. We processed summary statistics for 18 external GWAS summary statistics using the same pipeline, and estimated genetic correlations using the same LD reference panel. For the list of external GWAS, please see Supplementary Table 21.

Data Availability

Summary statistics for the meta-analysis have been deposited in the GWAS catalog, accession number GCST90809428.