A multi-ancestry genome-wide association study of dry eye disease and systemic comorbidities

Summary

Background

Dry eye disease (DED) affects up to 25% of the adult population, with chronic symptoms of pain and dryness often negatively impacting quality of life. Twin studies indicate a substantial heritability. However, the genetic architecture of DED is largely unknown.

Methods

We devised a case–control algorithm using a combination of ICD-9/10 codes and prescription records and validated its performance through manual chart review, identifying 132,637 DED cases and 352,201 controls across four ancestry groups in the Million Veteran Program biobank. We then performed multi-ancestry genome-wide and phenome-wide association scans. A replication analysis was performed in the Genetic Epidemiology Research on Adult Health and Aging biobank (up to 16,025 cases and 54,818 controls). Genetic correlations were assessed using linkage disequilibrium score regression and multivariate relationships were modelled using genomic structural equation modelling (genomicSEM).

Findings

We identify ten significant loci in nine susceptibility regions with largely consistent effects across ancestries, including loci linked to synapse maintenance (EFNA5, GRIA1, and SYNGAP1) and autoimmunity (BLK). Phenome-wide scans for genetic pleiotropy indicate substantial genetic correlations of DED with numerous comorbidities, including fibromyalgia (r_G = 0.69, 95% confidence interval, CI: [0.58, 0.81]), post-traumatic stress disorder (r_G = 0.58, 95% CI: [0.42, 0.75]), and Sjögren's disease (r_G = 0.40, 95% CI: [0.23, 0.57]). GenomicSEM identified a latent factor underlying DED and other chronic pain traits which accounted for 51% of the genetic variance of DED.

Interpretation

The observed genetic correlations are consistent with shared genetic liability between DED and chronic pain, autoimmune, and psychiatric traits, but do not establish causal direction.

Funding

This work was supported by the US Department of Veterans Affairs, the Department of Defense, the National Eye Institute/NIH, and Research to Prevent Blindness. Full funding details are provided in the Acknowledgements.

Research in context.

Evidence before this study

We searched PubMed from database inception to July 30, 2026, using the terms (“dry eye” OR “keratoconjunctivitis sicca”) AND (“genome-wide association” OR GWAS OR heritability OR polygenic OR “genetic correlation”). Twin studies supported a heritable component of dry eye disease (DED) and shared genetic influences between DED and chronic pain syndromes. Previous genetic studies examined self-reported DED in the Taiwan Biobank, neuropathic ocular pain, and selected relationships between DED and comorbidities or exposures. However, the genetic architecture of clinically recognised DED across ancestry groups and the extent to which associations with systemic comorbidities reflect shared genetic liability remained incompletely characterised.

Added value of this study

Using a validated electronic health record-based phenotype in the multi-ancestry Million Veteran Program, we characterised the genetic architecture of clinically recognised DED. We identified genome-wide significant susceptibility loci near genes with plausible roles in autoimmunity and neuronal signalling, including BLK, SYNGAP1, GRIA1, and EFNA5, with generally consistent effects across ancestry groups. Phenome-wide polygenic score analyses and genetic correlation analyses indicated extensive shared genetic liability between DED and chronic pain, psychiatric, autoimmune, cardiometabolic, and behavioural traits, including fibromyalgia, multisite chronic pain, post-traumatic stress disorder, and Sjögren's disease. Genomic structural equation modelling further supported a common genetic factor shared by DED and other chronic pain conditions, while also indicating genetic influences on DED that were not captured by the shared pain factor.

Implications of all the available evidence

The findings support investigating DED within a broader multisystem context, alongside its established ocular-surface manifestations. Our results suggest that the well-documented associations between DED and various systemic illnesses may partly reflect shared genetic factors. This interpretation supports caution regarding some presumed causal relationships, such as the link between digital screen use and DED, which may reflect shared genetic liability in addition to any direct causal effect. Elucidating the shared biology of DED and its systemic comorbidities could inform future therapeutic strategies.

Introduction

Dry eye disease (DED) is a heterogeneous condition defined by a compilation of patient-reported symptoms and physical signs seen on ocular examination.¹ Symptoms of DED include both vision- and pain-related symptoms, although in the majority of cases, pain-related symptoms predominate, with individuals using terms such as “dryness,” “aching,” “grittiness,” and “tenderness” to describe their pain.² Biological manifestations of DED observed as signs on ocular examination also vary and can include decreased tear production, epithelial disruption, and tear instability, in isolation or combination.³ Ocular surface inflammation and tear hyperosmolarity often accompany tear abnormalities.⁴ Adding to the complexity of disease, symptoms and signs of DED are often discordant⁵ and individuals may carry a DED diagnosis based on symptoms, signs, or a combination of both. Given this broad definition, it is not surprising that DED is a prevalent condition in the adult population, with an estimated frequency of 5%–30% worldwide.⁶ We previously demonstrated that DED is common in United States (US) veterans, with a prevalence of 19% in male and 22% in female veterans based on International Classification of Disease (ICD) coding and the use of at least one DED-associated medication.⁷

Several conditions and medications have been linked to DED.⁸ Common risk factors for DED include female sex, increasing age, medication use (e.g., antihistamines, antidepressants), and adverse environmental conditions such as low humidity and air pollution.⁶ In our previous study of US veterans, post-traumatic stress disorder (PTSD) and depression were strongly associated with a DED diagnosis.⁷ Other studies have reported associations between DED and pain conditions, including migraine,⁸ fibromyalgia, and pelvic pain.⁹ DED is also closely related to autoimmune disorders and is one of the defining symptoms of Sjögren's disease (SjD).¹⁰

Despite the complexity of defining DED, consistent associations with systemic comorbidities suggest shared heritable factors. Within-pair correlations for DED-related traits, were higher in monozygotic twins compared to dizygotic twins, suggesting an underlying genetic contribution.¹¹ Specifically, DED symptoms were estimated to have a heritability of 29% (95% CI: [18%, 40%]).¹¹ Twin studies in individuals with chronic pain conditions related to DED (e.g., migraine) also reported higher within-pair correlations in monozygotic compared to dizygotic twins.¹² Additionally, findings from the TwinsUK Adult Twin Registry suggested shared genetic factors between chronic widespread pain, pelvic pain, irritable bowel syndrome, and DED.¹³

The heterogeneity of DED presentation and a lack of uniformity in International Classification of Diseases (ICD) coding practices for DED have limited the application of biobank-scale genome-wide association studies (GWAS) to date. A Taiwan Biobank (TWB) analysis¹⁴ of self-reported DED in 14,185 cases and 25,927 controls reported 11 independent loci with suggestive significance (P < 1 × 10⁻⁵) though none reaching standard genome-wide significance (P < 5 × 10⁻⁸). A subsequent study found a genetic correlation of 0.19 between DED and depression in TWB.¹⁵ A South Florida study conducted a GWAS of 329 individuals with neuropathic ocular pain (NOP) symptoms (e.g., burning, wind sensitivity),¹⁶ and found one single nucleotide polymorphism (SNP), rs140293404, that reached genome-wide significance, suggesting the possibility of genetic variants predisposing an individual to the development of NOP. Similarly, large-scale GWAS on SjD have identified 22 susceptibility loci,^17,18 most of which are linked to immunity-related genes such as IRF5, STAT4, and BLK.

The heritability of clinically recognised DED in a large US population and its genetic relationships with systemic comorbidities remain incompletely characterised. While previous studies identified correlations between DED and various comorbidities,^8,9 it remains unclear whether there are shared genetic risk factors contributing to these associations. The Million Veteran Program (MVP)¹⁹ provides an opportunity to address this gap as it links genotypic information from a large cohort of US veterans to longitudinal medical records. The goal of this study is to examine the heritability of DED in a US veteran population and examine shared genetic factors between DED and DED-associated comorbidities. Given the prevalence of DED and its negative impact on quality of life, a better understanding of disease contributors, including those shared with non-ocular comorbidities, may ultimately inform precision-based treatment algorithms for individual patients. To accomplish this, we conducted a multi-ancestry discovery GWAS of clinically recognised DED in MVP and evaluated replication of lead loci in the independent Genetic Epidemiology Research on Adult Health and Aging (GERA) cohort.

Methods

Ethics

MVP is an ongoing prospective cohort and mega-biobank within the Veterans Health Administration (VHA) enrolling US veterans who are active users of VHA care, have available electronic health records, and provide written informed consent¹⁹ (https://www.research.va.gov/mvp/). Enrolment began on January 5, 2011, at VA facilities in Boston, Massachusetts, and West Haven, Connecticut, with subsequent expansion to recruitment centres nationwide. For the present analysis, the final date of EMR follow-up or DED case–control ascertainment was September 30, 2022. At the time of this study, MVP included genetic data for 658,582 veterans, which constituted the source population for phenotype assignment and downstream genetic analyses. Local chart reviews were conducted at the eye clinics of three VA Medical Centres (VAMCs): VA Miami Healthcare System (Miami, FL), VA Northeast Ohio Healthcare System (Cleveland, OH), and Providence VA Medical Center (Providence, RI). This study adhered to the tenets of the Declaration of Helsinki, as revised in 2024. VA-based study procedures were approved by the VA Office of Research & Development Central Institutional Review Board (VA CIRB 18–41). The GERA cohort was approved by the Institutional Review Board of the Kaiser Foundation Research Institute under the Research Program on Genes, Environment, and Health Grand Opportunity Project (IRB CN-09CScha-06-H), and all GERA participants provided written informed consent.

Case-control algorithm

We devised a case–control algorithm based on ICD-9/10 codes and prescription records. Inclusion criteria for cases included at least two independent diagnoses of DED on different days, based on ICD-9 codes (370.33, 375.15) or ICD-10 codes (H04.121, H04.122, H04.123, H04.129, H16.221, H16.222, H16.223, H16.229, and M35.01) and at least one prescription for a medication used for DED (as listed in the drug database) (Supplementary Table S1). Individuals were included as a control if they had none of the diagnoses of DED listed above and no prescription for any of the DED medications in Supplementary Table S1. To test the algorithm for validity, it was applied to identify cases and controls at Miami, Providence, and Cleveland VAMCs, followed by manual validation of case–control status through chart review at each site. Once satisfied with the overall multi-site positive predictive value (PPV) and negative predictive value (NPV), the validated algorithm was applied to the entire MVP biobank sample to obtain the population for our study. We identified 491,962 individuals overall, with 134,249 DED cases and 357,713 controls (Table 1). This phenotype-defined cohort served as the starting population for defining the subsequent GWAS analytic sets.

Table 1.

Demographics in the study population.

Demographics	Total population (n = 491,962)	Cases	Controls	Odds ratio (95% CI)
		(DED) (N = 134,249)	(No DED) (N = 357,713)
Age, years, mean ± SD	67.6 ± 14.62	72.27 ± 11.16	65.83 ± 15.36	1.05 (1.05–1.05)
Sex, N (%)
Male	447,272 (90.9)	118,750 (88.5)	328,522 (91.8)	-
Female	44,690 (9.1)	15,499 (11.5)	29,191 (8.2)	2.38 (2.33–2.44)
Ancestry, N (%)
EUR	345,672 (70.3)	87,444 (65.1)	258,228 (72.2)	-
AFR	89,069 (18.1)	30,734 (22.9)	58,335 (16.3)	1.99 (1.96–2.02)
AMR	44,804 (9.1)	12,940 (9.6)	31,864 (8.9)	1.77 (1.72–1.81)
EAS	5,293 (1.1)	1,519 (1.1)	3,774 (1.1)	1.88 (1.76–2.00)
SAS	415 (0.1)	131 (0.1)	284 (0.1)	3.20 (2.57–3.97)
Other	6,709 (1.4)	1,481 (1.1)	5,228 (1.5)	1.18 (1.11–1.26)
Body mass index, mean ± SD	29.70 ± 5.82	30.32 ± 5.98	29.46 ± 5.74	1.04 (1.04–1.04)
Smoking status, N (%)
Never	121,958 (24.8)	34,091 (25.4)	87,867 (24.6)	-
Current	114,618 (23.3)	28,932 (21.6)	85,686 (24.0)	1.06 (1.04–1.08)
Prior	188,978 (38.4)	60,693 (45.2)	128,285 (35.9)	1.15 (1.13–1.17)
Missing	66,408 (13.5)	10,533 (7.8)	55,875 (15.6)	0.57 (0.55–0.58)

N: sample size; DED: dry eye disease; CI: 95% confidence interval; SD: standard deviation; EUR: European ancestry; AFR: African ancestry; AMR: Admixed American ancestry; EAS: East Asian ancestry; SAS: South Asian ancestry.

Odds ratios and 95% confidence intervals were estimated using multivariable logistic regression. Models were adjusted for age, sex, and genetic ancestry where applicable, excluding the predictor of interest from the adjustment set. Odds ratios for age and BMI are per 1-unit increase. BMI was missing for 3771 participants overall, including 213 DED cases and 3558 controls; BMI means, SDs, and odds ratios were calculated among participants with non-missing BMI.

Comorbidities

Comorbidities and medication use were gathered from patient-reported survey data (from Lifestyle and Baseline MVP surveys) and EMR searches for Charlson Comorbidity Index (CCI) variables.

Genetic ancestry

Participants in MVP were classified based on similarity to one of five genetic ancestry clusters using external reference panels²⁰: European (EUR), African (AFR), Admixed American (AMR), East Asian (EAS), and South Asian (SAS). The SAS ancestry cluster was excluded from GWAS analyses due to the small sample size (<200 cases). Samples that did not have a high degree of similarity to any cluster (“Other”) were also excluded from GWAS analyses. Within each ancestry cluster, we performed principal components analysis (PCA) to assess population substructure and generate covariates for genetic association analyses. PCA was performed using PLINK 1.9²¹ on linkage disequilibrium (LD) pruned SNPs (“--indep-pairwise 50 10 0.1”) in a maximal set of unrelated samples, defined as excluding pairs of individuals related within two degrees, and the full sample, including related individuals, was projected onto the resulting eigenvectors. Self-reported race and ethnicity were not used to define ancestry groups or as covariates in the present analyses; the ancestry labels refer to genetically inferred similarity clusters and should not be interpreted as self-identified race or ethnicity.

Sex/gender

Genetically inferred sex, based on sex chromosome karyotype estimated from genotyping array intensities, was used as the sex variable in all analyses. Self-reported gender was not used in the present analyses.

Association of comorbidities and medication use with DED

A multiple logistic regression model was performed for each comorbidity and medication, with DED as the dependent variable, using age, sex, and genetic ancestry as covariates. Individuals with missing comorbidity data (N = 7083) were excluded; complete-case analysis was used for each model.

Associations of polygenic scores (PGSs) with DED

Phenome-wide polygenic score files were obtained from the European Molecular Biology Laboratory's European Bioinformatics Institute (EMBL-EBI) PGS Catalog.²² The database of PGS scores was downloaded in September 2022. PGSs using MVP in the training data were excluded. After excluding PGSs that used MVP in the training data, 2651 PGSs were available for association testing. All subjects in MVP were scored across all available PGSs using the BCFtools/score plugin (https://github.com/freeseek/score) of BCFtools²³ v1.17. PGSs were then loaded into the dosage format field of VCFs readable by SAIGE for association testing. To determine the pleiotropy of genetic liability to traits on DED, we performed a scan of association between PGSs and DED case–control status using logistic regression models implemented in SAIGE²⁴ v1.3.0, adjusting for sex, age, mean-centred age-squared, and ten ancestry-specific principal components. For visualisation, PGS Catalog phenotype labels were classified into broad disease-system categories based on the PheWAS framework²⁵ using the GPT-4 Turbo²⁶ large language model (snapshot gpt-4-0125-preview). The resulting assignments were manually reviewed for face validity. The exact prompt, category labels, and classification script are provided in the Supplementary Information.

Statistics

Detailed MVP genotyping laboratory procedures, allele calling, sample-level and variant-level quality control, genotyping success rates, and array performance metrics have been described previously.²⁷ In brief, MVP samples were genotyped on the MVP 1.0 Axiom array. An ancestry-aware Hardy–Weinberg Equilibrium filter was applied (P < 1 × 10⁻⁶ within each major ancestry group in MVP) prior to imputation. Samples were pre-phased using SHAPEIT4²⁸ and imputed to the TOPMed imputation panel version r2 (N = 97,256 sequences) using Minimac4.²⁹ GWAS analyses were performed on ancestry-stratified subsets in MVP using SAIGE²⁴ v1.3.0, adjusting for sex, age, mean-centred age-squared, and ten ancestry-specific principal components. Related individuals were retained in the GWAS analyses, and genetic relatedness was accounted for using SAIGE logistic mixed models, which include a random genetic effect with covariance structure specified by the genetic relatedness matrix. To ensure accurate effect size estimation, Firth approximation was applied to SNPs with P < 0.05. Association scans were performed on well-imputed SNPs (INFO >0.5) using an ancestry-specific minor allele frequency (MAF) cutoff of ≥0.5% and a minimum minor allele count (MAC) cutoff of 20. Additional sensitivity GWAS analyses were performed in EUR following the same methods as the primary analyses: exclusion of all individuals with a clinical history of SjD (at least one diagnosis code from ICD-9 710.2 or ICD-10 M35.0∗), and sex-stratified GWAS (Supplementary Table S2). No randomisation or blinding was performed because this was an observational biobank study. Sample size was determined by available MVP participants meeting case–control and genetic QC criteria.

GWAS meta-analysis

We performed inverse variance-weighted fixed-effect meta-analyses of the ancestry-stratified GWAS summary statistics. Each set of summary statistics was converted into GWAS-VCFs³⁰ using the BCFtools/munge plug-in (https://github.com/freeseek/score) of BCFtools²³ v1.17, followed by meta-analysis using the BCFtools/metal plug-in (https://github.com/freeseek/score).

Clumping

We performed clumping to identify independent loci within broad genome-wide significant susceptibility regions. Reference samples were extracted from the EUR superpopulation of the 1000 Genomes Phase 3. We performed clumping using PLINK 1.9²¹ with an r² < 0.2 threshold and a 2 Mb window. We further validated that the independent loci identified using this method had r² < 0.2 in all ancestries analysed (EUR, AFR, AMR, and EAS).

Pleiotropy analysis

We used LDtrait³¹ to perform a scan for pleiotropic associations with each DED locus reported in the GWAS Catalog.³² Associations with P < 5 × 10⁻⁸, LD r² > 0.1, and located within 500 kilobases of the DED locus were included. The GWAS Catalog data were retrieved in February 2025.

Replication analysis

Replication was performed in the multiethnic GERA cohort. The GERA cohort contains genome-wide genotype, clinical, and demographic data of over 110,000 adult members of the Kaiser Permanente Northern California integrated healthcare system.^33,34 GERA samples were genotyped at over 665,000 genetic markers on four ethnic-specific Affymetrix Axiom arrays optimised for European, Latino, East Asian, and African American individuals.^35,36 Imputation was performed by array with Minimac3 v2.0.1²⁹, using either the Haplotype Reference Consortium (HRC) reference panel³⁷ or the 1000 Genomes Project Phase III reference panel.³⁸ In GERA, two phenotypes were defined: a narrow phenotype (based on DED medication prescription and/or procedure records as listed in Supplementary Table S3), which was most similar to the phenotype used in MVP; and a broad phenotype (at least 2 DED diagnoses based on ICD-10 code H04.12 or ICD-9 code 375.15, but no required medication prescription or procedure records). Controls had no DED diagnoses. In GERA, while 3317 DED cases and 54,516 controls were identified using the narrow definition, 16,025 DED cases and 54,818 controls were identified using the broad definition. Statistical power was calculated using the Genetic Association Study (GAS) Power Calculator.³⁹

Linkage disequilibrium score regression (LDSC)

Heritability for DED and pairwise genetic correlations (r_G) between DED and other traits were computed using LDSC⁴⁰ v1.0.1. Prior to computing r_G, all summary statistics were quality-controlled and alleles were harmonised to the GRCh38 reference genome using the BCFtools/liftover⁴¹ and BCFtools/munge plug-ins. European-ancestry LD scores provided by the LDSC software were used, which exclude the major histocompatibility complex (MHC) region as standard practice. To further characterise the biological relevance of the DED GWAS, we performed tissue-specific heritability enrichment using LDSC-SEG⁴² across 205 tissue annotations via the Complex-Traits Genetics Virtual Lab⁴³ (CTG-VL; vl.genoma.io).

Genomic structural equation modelling (GenomicSEM)

We applied genomicSEM⁴⁴ to examine the genetic overlap between DED and other related disorders. genomicSEM estimates the genetic variance-covariance matrix using LDSC from GWAS summary statistics and is robust to sample overlap. We sought to replicate the multivariate analysis of DED and other pain-related conditions performed in a prior twin study.¹³ We performed a single-factor confirmatory factor analysis using GWAS summary statistics for DED, fibromyalgia, irritable bowel syndrome, multisite chronic pain, and migraine. We evaluated model fit using the model chi-square statistic, fit indices (Comparative Fit Index, or CFI; Standardised Root Mean Square Residual, or SRMR), and standardised factor loadings. We considered CFI >0.90, SRMR <0.05, and factor loadings >0.50 to indicate adequate model fit.

Role of funders

The funders of this study had no role in the design, data collection, data analysis, data interpretation, or writing of this manuscript.

Results

Algorithm validation

A summary of the overall experimental design is provided in Fig. 1. Validation of the DED algorithm at the Miami VA hospital noted a 100% (41/41) PPV and a 96% (49/51) NPV for differentiating DED cases from controls. This algorithm was then independently validated at two other VA sites (Providence and Cleveland), with a combined PPV of 100% (100/100) and NPV of 98% (98/100). Validation was based on the documentation of DED symptoms, signs, or diagnosis in the clinical note. Across all sites, the PPV was 100% and the NPV was 97%.

Fig. 1.

Overall study design. We developed and validated a dry eye disease (DED) phenotyping algorithm based on diagnosis codes and prescription records and applied it to genotyped participants in the Million Veteran Program (MVP). Manual chart review was used to estimate algorithm performance across three Veterans Affairs Medical Centres. The phenotype-defined MVP cohort included 134,249 DED cases and 357,713 controls, from which ancestry-specific GWAS analytic sets were defined. Participants were assigned to genetic ancestry groups by similarity to external reference panels; four groups had sufficient sample size for ancestry-stratified GWAS analyses: EUR, AFR, AMR, and EAS. Polygenic score association scans and GWAS were performed within ancestry groups, and GWAS summary statistics were combined in a multi-ancestry meta-analysis. Replication of the ten genome-wide significant loci was performed in the Genetic Epidemiology Research on Adult Health and Aging (GERA) cohort. The GERA sample size shown corresponds to the broad DED phenotype. Heritability, genetic correlation, and genomic structural equation modelling analyses were performed using EUR GWAS summary statistics. VAMC: Veterans Affairs Medical Center; EUR: European ancestry; AFR: African ancestry; AMR: Admixed American ancestry; EAS: East Asian ancestry; PGS Catalog: Polygenic Score Catalog; GWAS: genome-wide association study.

MVP study population

The total population comprised 658,582 genotyped participants, of which 491,962 individuals (134,249 DED cases and 357,713 controls) met our case–control inclusion criteria (Fig. 1). Participants were classified by genetic similarity to five major continental ancestry clusters: European (EUR), African (AFR), Admixed American (AMR), East Asian (EAS), and South Asian (SAS).²⁰ Four similarity clusters (EUR, AFR, AMR, and EAS) had at least 200 DED cases and thus formed the analytic set for ancestry-stratified genetic analyses (combined 132,637 cases and 352,201 controls). The mean age of the population was 67.6 ± 14.6 years, with the majority having genetically inferred male sex (90.9%) and EUR ancestry (70.3%). Of the total population, 27.3% were classified as a case, meeting the inclusion criteria for a diagnosis of DED. The cases were older, less likely to be male, and more likely to be non-EUR in comparison to the controls (Table 1). Additional comorbidities and medication use in the total population are listed in Supplementary Table S4. Overall, DED cases had poorer health compared to controls, with a higher percentage of comorbidities and medication use in all domains, including neurologic, rheumatologic, and psychiatric.

Phenome-wide polygenic score scan

DED is associated with numerous comorbidities, but it is unclear to what extent these associations can be explained by shared genetic liability. To comprehensively assess the relationship between DED and inherited liability to comorbid traits, we performed a hypothesis-free phenome-wide association scan between DED case–control status and all available polygenic scores in the PGS Catalog. Because available polygenic scores vary substantially in discovery GWAS sample size, ancestry composition, trait representation, and predictive performance, these analyses should be interpreted as an exploratory scan of currently available PGS resources. As presently most polygenic scores have been developed from GWAS in the UK Biobank (UKB) and other European biobanks, we initially focused on the association of PGSs with our EUR DED cohort. Any scores that used MVP as a training set for the base GWAS were excluded. In EUR participants, 569 polygenic scores, representing 194 unique phenotypes mapping to the Experimental Factor Ontology, were significant at a strict Bonferroni cutoff based on the number of scores (P < 1.89 × 10⁻⁵). The single most strongly associated PGS with DED case–control status was self-rated health (PGS002218); each standard deviation increase in the PGS was associated with an odds ratio (OR) of 1.14 (95% CI: [1.13, 1.15]). Numerous other PGSs representing genetic liability to diseases across nearly all organ systems were significantly associated with DED (Fig. 2; Supplementary Table S5).

Fig. 2.

Phenome-wide scan of association between disease-related polygenic scores and dry eye disease case–control status. Each polygenic score (PGS) in the PGS Catalog was tested for association with DED case–control status in MVP. For visualisation, only PGSs whose phenotype labels were assigned to one of the 17 PheWAS-based disease-system categories are shown. Results shown are for EUR participants only, comprising 87,444 DED cases and 258,228 controls. Associations were tested using logistic regression models implemented in SAIGE, adjusted for sex, age, mean-centred age-squared, and ten ancestry-specific principal components. P-values are two-sided and were compared with a Bonferroni significance threshold based on the number of tested scores. Full summary statistics are provided in Supplementary Table S5. ADHD: attention deficit hyperactivity disorder; COPD: chronic obstructive pulmonary disease; GERD: gastro-oesophageal reflux disease; T2D: type 2 diabetes.

We then assessed the association of the polygenic scores in other ancestries. In general, the cross-ancestry predictivity of polygenic scores, also known as transferability, is reduced as a function of genetic distance between the training and target cohorts.⁴⁵ The combination of smaller cohort sizes in our non-EUR sets and attenuated cross-ancestry PGS transferability greatly reduces statistical power. Therefore, lack of association in a non-EUR ancestry group should not be interpreted as evidence of absence, nor should differences in association strength across ancestries be interpreted as direct evidence of ancestry-specific biology. Of the 569 Bonferroni-significant scores in EUR, 173 scores (representing 78 unique traits) were nominally significant (P < 0.05) in AFR with directionally consistent effects, 264 scores (106 traits) were nominally significant in AMR, and 14 scores (10 traits) were nominally significant in EAS (Supplementary Table S5). Association z-scores were highly correlated across EUR, AFR, and AMR ancestries (Supplementary Figure S1). Six scores (five traits) were Bonferroni-significant in EUR and significant at P < 0.05 in AFR, AMR, and EAS, with consistent effect direction: major depressive disorder (MDD) (PGS000907), number of non-cancer illnesses (a survey-based phenotype from UKB; PGS001004), abnormal appetite (PGS002098), neuroticism (PGS002213, PGS001996), and peripheral nerve disorders (PGS002039). Overall, these results suggest that several DED-associated polygenic liabilities, particularly involving psychiatric, general health, and pain-related traits, show supportive evidence across ancestry groups.

GWAS for DED

We performed a GWAS in each ancestry in MVP, followed by a multi-ancestry meta-analysis. The individual population GWAS and meta-analysis all had genomic inflation factor λ ≤ 1.18 (λ₁₀₀₀ ≤ 1.001), indicating minimal systematic inflation of the test statistic (Supplementary Figure S2; Supplementary Table S2). In our multi-ancestry meta-analysis, ten independent loci within nine susceptibility regions reached genome-wide significance (Fig. 3; Supplementary Figure S3; Table 2). None of the loci were located within 1 Mb of any of the suggestively significant loci previously reported in the Taiwan Biobank.¹⁴

Fig. 3.

Manhattan plot of GWAS results for dry eye disease. The multi-ancestry meta-analysis of four ancestries represented in the Million Veteran Program is shown (total of 132,637 cases and 352,201 controls). Within each ancestry, variant associations with DED case–control status were tested using SAIGE logistic mixed models adjusted for sex, age, mean-centred age-squared, and ten ancestry-specific principal components. Ancestry-specific GWAS summary statistics were combined using inverse variance-weighted fixed-effect meta-analysis. The x-axis shows chromosomal position and the y-axis shows −log10(P), where P-values are two-sided association P-values from the multi-ancestry meta-analysis. The red line denotes the genome-wide significance threshold (P < 5 × 10⁻⁸). Genome-wide significant loci at each of the nine susceptibility regions are highlighted and labelled with the nearest gene.

Table 2.

Genome-wide significant loci in the multi-ancestry meta-analysis of dry eye disease (DED).

rsID	Chr:pos	Nearest gene	EA/NEA	EAF	N cases	N	OR [L95, U95]	P-value	Effect direction	I2	Het. P-value
rs4682749	3:44,697,023	ZNF502	G/C	0.39	132,637	484,838	0.97 [0.96, 0.98]	9.95 × 10−9	---+	28.1	0.24
rs13107325	4:102,267,552	SLC39A8	T/C	0.08	131,118	479,545	1.07 [1.04, 1.09]	1.24 × 10−10	+++?	0.0	0.90
rs2411044	5:107,737,792	EFNA5	A/G	0.17	132,637	484,838	0.96 [0.95, 0.98]	2.77 × 10−8	----	0.0	0.68
rs7731539	5:153,732,101	GRIA1	G/C	0.48	132,637	484,838	0.97 [0.96, 0.98]	2.84 × 10−8	----	22.0	0.28
rs34025038	5:153,801,313	GRIA1	T/C	0.12	132,637	484,838	1.05 [1.03, 1.06]	9.81 × 10−9	++++	0.0	0.75
rs191549504	6:33,443,896	SYNGAP1	C/T	0.01	88,963	350,965	1.16 [1.11, 1.21]	4.18 × 10−10	+??-	0.0	0.50
rs2618471	8:11,501,511	BLK	G/C	0.35	132,637	484,838	1.03 [1.02, 1.04]	2.91 × 10−8	+++-	0.0	0.45
rs12779865	10:21,540,233	MLLT10	C/T	0.31	132,637	484,838	1.04 [1.03, 1.05]	1.58 × 10−13	+++-	35.2	0.20
rs77465292	11:47,586,360	FAM180B	C/T	0.13	132,637	484,838	0.96 [0.94, 0.97]	6.30 × 10−9	---+	65.8	0.03
rs2872878	19:19,369,712	GATAD2A	C/G	0.59	132,637	484,838	0.97 [0.96, 0.98]	4.77 × 10−9	----	0.0	0.65

Chr: chromosome; Pos: position; EA: effect allele; NEA: non-effect allele; EAF: effect allele frequency; N cases: number of cases; N: total sample size; OR [L95, U95]: odds ratio with lower and upper bounds of the 95% confidence interval; Effect Direction: SNP effect direction in the EUR, AFR, AMR, and EAS cohorts, respectively (“?” indicates the variant did not meet the allele frequency cutoff of 0.5% in that ancestry and was not included in the meta-analysis); I²: percentage of variation due to heterogeneity; Het. P-value: P-value of the Cochran's Q heterogeneity test.

We identified 10 independent loci within 9 susceptibility regions. Genomic coordinates correspond to the GRCh38 assembly. Ancestry-stratified summary statistics for each locus are provided in Supplementary Table S6.

Of the ten loci, five were genome-wide significant in EUR; otherwise, discovery relied on the added statistical power from combining multiple ancestries (Supplementary Table S6). Effect sizes were largely consistent across the four ancestry groups; all loci had heterogeneity P > 0.05 except for one (Table 2). This locus, rs77465292 (near FAM180B), had a stronger effect in AFR (OR = 0.89, 95% CI: [0.84,0.95], P = 4.5 × 10⁻⁴) than in EUR (OR = 0.96, 95% CI: [0.94,0.97], P = 2.3 × 10⁻⁷). The three largest populations (EUR, AFR, and AMR) had consistent effect direction across all loci.

All ten loci exhibited pleiotropic associations with numerous other traits, including comorbidities linked to DED (Supplementary Table S7). For instance, eight loci were linked with body mass index (BMI), eight with insomnia, four with depression or MDD, and two with multisite chronic pain. One locus, rs2618471 (near BLK), had many autoimmune disease pleiotropic associations, including SjD, systemic sclerosis, systemic lupus erythematosus, and rheumatoid arthritis; this was the only locus that overlapped with SjD. Interestingly, rs77465292 (near FAM180B) was linked with central corneal thickness, corneal resistance factor, and primary open-angle glaucoma (POAG), but the POAG risk allele was protective for DED.

Replication of genetic results

We sought replication of our GWAS loci in the independent multiethnic GERA cohort. Demographics of the GERA cohort are found in Supplementary Table S8. Of the ten loci that reached genome-wide significance in MVP, only one (rs12779865) was powered at 80% probability to identify P < 0.05 in the GERA broad phenotype and none in the narrow phenotype. rs12779865 showed nominal, directionally consistent support in the GERA broad-phenotype analysis (OR = 1.03, 95% CI: [1.00, 1.06], P = 0.026; Supplementary Table S9). Overall, seven of the nine loci genotyped in GERA had consistent effect direction with MVP (Supplementary Figure S4).

Heritability

We estimated the heritability attributable to additive common variation (SNP-h²) using LDSC with the DED EUR summary statistics. The liability scale SNP-h² was 6.0% (standard error 0.37%), assuming a population prevalence of 25%, for a highly significant heritability z-score of 16.1. The LD intercept was 1.04 (s.e. 0.0075), corresponding to an attenuation ratio of 0.15 (s.e. 0.028), consistent with limited residual inflation. Overall, these results indicate that our GWAS is well-calibrated and captures significant heritable signal, with most of the inflation in test statistics attributable to polygenicity rather than population stratification or other confounding. In the sex-stratified sensitivity analysis, we derived similar SNP-h² estimates for EUR males (6.1%, s.e. 0.40%) and females (8.7%, s.e. 2.9%), assuming a population prevalence of 25% for both. Male and female DED were highly genetically correlated (r_G = 0.71; 95% CI: [0.41, 1.01]).

Genetic correlation

Based on the results of our PGS association discovery scan, we selected traits that either reached Bonferroni-adjusted significance or were related to significant traits to perform genome-wide genetic correlation analyses. For each trait, we used the largest consortium GWAS where summary statistics were publicly available. Otherwise, we used GWAS from FinnGen,⁴⁶ the Pan-UK Biobank project⁴⁷ (Pan-UKB), or, where traits were harmonised across the two biobanks, a meta-analysis of FinnGen and Pan-UKB generated by the FinnGen consortium; details are provided in Supplementary Table S10. For this analysis, we focused on the EUR DED summary statistics. All summary statistics for the other traits were based on European ancestry GWAS, and none included data from MVP. Because these traits were established in the discovery scan, we considered P < 0.05 to be statistically significant evidence of genetic correlation.

Of the 54 traits selected for genetic correlation analysis, 52 demonstrated statistically significant genetic correlations, and 46 had P < 0.001, consistent with the PGS associations (Fig. 4; Supplementary Table S11). The trait with the highest genetic correlation with DED was fibromyalgia (r_G = 0.69; 95% CI: [0.58,0.81]), followed by multisite chronic pain (r_G = 0.64, 95% CI: [0.58, 0.70]). Other chronic pain-related traits showed substantial genetic correlation with DED, including gastro-oesophageal reflux (r_G = 0.53, 95% CI: [0.45, 0.60]), asthma (r_G = 0.45, 95% CI: [0.38, 0.53]), irritable bowel syndrome (r_G = 0.46, 95% CI: [0.36, 0.56]), and osteoarthritis (r_G = 0.40, 95% CI: [0.33, 0.47]). Notably, a moderate genetic correlation was also found with endometriosis (r_G = 0.32, 95% CI: [0.24, 0.41]), despite the cohort being 90% male.

Fig. 4.

Genetic correlations of selected traits with dry eye disease. Genetic correlations between DED and selected traits were estimated using linkage disequilibrium score regression (LDSC) with EUR DED GWAS summary statistics and publicly available EUR-ancestry GWAS summary statistics for comparison traits. Points represent genetic correlation estimates (r_G), and horizontal error bars represent 95% confidence intervals calculated from LDSC standard errors. P-values are two-sided LDSC z-test P-values testing whether r_G differs from zero. Stars indicate P < 0.001, unadjusted for the number of tested traits. Traits were selected because they were significant, or related to significant traits, in the phenome-wide PGS association scan. GERD: gastro-oesophageal reflux disease; COPD: chronic obstructive pulmonary disease; PTSD: post-traumatic stress disorder; ADHD: attention deficit hyperactivity disorder; HDL cholesterol: high-density lipoprotein cholesterol; 95% c.i.: 95% confidence interval; PGS: polygenic score.

Psychiatric and neurological traits also showed substantial genetic correlation with DED. Consistent with our earlier work,⁷ the psychiatric trait with the strongest correlation was PTSD (r_G = 0.58, 95% CI: [0.42, 0.75]), followed by attention-deficit/hyperactivity disorder (ADHD) (r_G = 0.45, 95% CI: [0.39, 0.52]), MDD (r_G = 0.36, 95% CI: [0.30, 0.42]), and broadly defined anxiety (r_G = 0.33, 95% CI: [0.25, 0.41]). Bipolar disorder (r_G = 0.10, 95% CI: [0.04, 0.16]) and schizophrenia (r_G = 0.06, 95% CI: [0.01, 0.12]) showed relatively attenuated genetic correlations with DED, consistent with known interrelationships among disorders on the psychosis-affective spectrum.⁴⁸ Neurological traits of fatigue (r_G = 0.55, 95% CI: [0.47, 0.62]) and insomnia (r_G = 0.42, 95% CI: [0.34, 0.49]) were also correlated. Finally, neuroticism, a personality trait, was significantly correlated with DED (r_G = 0.24, 95% CI: [0.14, 0.34]).

Autoimmune traits represented another major nexus of genetic correlation with DED. Of autoimmune traits tested, SjD had the highest genetic correlation (r_G = 0.40, 95% CI: [0.23, 0.57]), consistent with dry eyes being a key symptom of SjD. In a sensitivity analysis where individuals with any history of SjD diagnosis were excluded from cases and controls, the genetic correlation with SjD remained virtually the same (r_G = 0.36, 95% CI: [0.19, 0.53]). Other autoimmune traits with moderate genetic correlations included allergic rhinitis (r_G = 0.38, 95% CI: [0.24, 0.52]) and juvenile idiopathic arthritis (r_G = 0.31, 95% CI: [0.16, 0.46]). A GWAS based on a diagnosis of any autoimmune disease had r_G = 0.21 (95% CI: [0.14, 0.29]). Notably, these genetic correlations were found despite excluding the MHC region from the analysis, following standard practice.

Cardiometabolic traits also featured prominently. Amongst cardiometabolic traits, sleep apnoea had the strongest genetic correlation (r_G = 0.42, 95% CI: [0.33, 0.50]), followed by type 2 diabetes (r_G = 0.35, 95% CI: [0.29, 0.41]), which has reported inflammatory and autoimmune facets. We also found moderate genetic correlations with triglycerides (r_G = 0.18, 95% CI: [0.09, 0.26]) and HDL cholesterol (r_G = −0.25, 95% CI: [−0.31, −0.19]).

Finally, several behavioural traits were genetically correlated with DED. Poor self-rated health was the most correlated behavioural trait (r_G = 0.55, 95% CI: [0.50, 0.61]). Consistent with ADHD, externalising behaviour⁴⁹ was a major theme, featuring genetic correlations with earlier tobacco use (r_G = −0.37, 95% CI: [−0.44, −0.30]), earlier sexual debut (r_G = −0.39, 95% CI: [−0.45, −0.33]), and lower educational attainment (r_G = −0.37, 95% CI: [−0.42, −0.32]). Whereas the PGS for general screen usage (computer or TV) was associated with DED, here we assessed genetic correlation with computer usage and TV usage separately. While time spent watching TV (r_G = 0.26, 95% CI: [0.19, 0.32]) and gaming (r_G = 0.15, 95% CI: [0.09, 0.22]) were correlated, general computer usage was not (r_G = 0.00, 95% CI: [−0.06, 0.06]). Interestingly, alcohol use measured in drinks per week was negatively correlated with DED (r_G = −0.09, 95% CI: [−0.15, −0.03]), potentially reflecting the complex genetic relationship between alcohol use and MDD and other psychiatric traits.⁵⁰ Finally, we observed a genetic correlation with increased water intake (r_G = 0.12, 95% CI: [0.05, 0.19]), consistent with observational associations.⁵¹

GenomicSEM

Fig. 5 presents the results from genomicSEM analyses of DED, fibromyalgia, irritable bowel syndrome, multisite chronic pain, and migraine. Overall, there were moderate to large genetic correlations across each of the phenotypes (Fig. 5A). Fig. 5B presents a single factor model, which showed good fit to the data (chi-square = 3.91, degrees of freedom (df) = 5, P = 0.56; comparative fit index (CFI) = 0.99; and standardised root mean square residual (SRMR) = 0.02). All of the indicators displayed moderate-to-large loadings (i.e., ≥0.5) on the single factor, and the residual variance of DED was 0.49. The latent “pain” common factor showed relatively strong genetic correlations with other autoimmune disorders, psychiatric conditions such as depression or PTSD, and cardiometabolic outcomes, among others (Fig. 5C). Compared to the bivariate genetic correlations between DED and each of these phenotypes, the estimates often became stronger and more precise (Supplementary Table S12). Lastly, LDSC-SEG heritability enrichment results showed nominally significant enrichment of the common factor in brain-related tissues, whereas the DED-specific residual factor showed nominal enrichment in ocular and endocrine-related annotations, including eye, pituitary, and thyroid-related tissues (Supplementary Figure S5; Supplementary Table S13). This pattern is consistent with a model in which the common factor captures centralised pain-related biology, while DED-specific residual genetic variance may reflect more ocular, peripheral, or neuroendocrine components of DED.

Fig. 5.

Genomic structural equation modelling of DED identifies shared pain risk. (A) Heatmap of pairwise genetic correlations between DED and representative chronic pain traits. The SNP heritability (SNP-h²) is represented along the diagonal. (B) The genomicSEM model used to derive a latent factor representing the shared genetic risk underlying DED and other chronic pain traits. The arrows pointing from the shared latent factor to each trait represent the factor loading, and the arrows pointing from u to each trait represent the residual variance. Model fit indices are provided below the model diagram. (C) Forest plot of genetic correlations of the latent factor with DED-linked traits of interest. Analyses used European-ancestry GWAS summary statistics for DED from MVP (87,444 cases and 258,228 controls); fibromyalgia and migraine from meta-analyses of FinnGen and UK Biobank data (4781 cases and 594,557 controls, and 46,682 cases and 751,137 controls, respectively); irritable bowel syndrome from a meta-analysis of UK Biobank and cohorts participating in the Bellygenes initiative (53,400 cases and 433,201 controls); and multisite chronic pain from UK Biobank (N = 387,649). Pairwise genetic correlations and genetic covariance matrices were estimated using LDSC, and the single-factor confirmatory model was fitted using genomicSEM. Full data sources and sample sizes for the traits shown in panel C are provided in Supplementary Table S10. IBS: irritable bowel syndrome; MCP: multisite chronic pain; SEM: structural equation modelling; u: residual variance; df: degrees of freedom; AIC: Akaike information criterion; CFI: comparative fit index; SRMR: standardised root mean square residual; rG: genetic correlation; 95% CI: 95% confidence interval; BMI: body mass index; COPD: chronic obstructive pulmonary disease; GERD: gastro-oesophageal reflux disease; ADHD: attention deficit hyperactivity disorder; PTSD: post-traumatic stress disorder.

Discussion

This multi-ancestry GWAS meta-analysis of 132,637 DED cases and 352,201 controls in MVP identified ten independent genome-wide significant loci in nine susceptibility regions. All loci had widespread multi-systemic pleiotropy with other traits, with generally consistent direction of effects with known DED-linked comorbidities. Several loci were within or near genes linked to plausible DED disease mechanisms, such as autoimmunity (BLK) and neurotransmitter signalling and synaptic plasticity (SYNGAP1, GRIA1, and EFNA5). Notably, all loci had relatively small effect sizes; the largest effect was OR = 1.16 at rs191549504 (SYNGAP1), a missense mutation with a frequency of just 1% in EUR (Supplementary Figure S6). These observations are consistent with DED being a highly polygenic trait, with many loci having at most a minor contribution to disease liability. Additionally, this illustrates the importance of large sample sizes, such as those provided by MVP, for reliably detecting individual loci associated with DED. In that sense, the genetic architecture of DED appears more similar to psychiatric traits, such as depression,⁵² than other complex ocular disease traits, such as age-related macular degeneration⁵³ or glaucoma,⁵⁴ which are partially characterised by common loci with large effect sizes.

Our replication analysis provided nominal support for the most significant locus from our GWAS (at MLLT10) but was underpowered to confirm the remaining loci. Thus, the lack of replication for several loci should not be interpreted as evidence against their validity, but rather as reflecting the limited power of currently available independent DED datasets to detect variants of small effect. Future efforts should pool harmonised DED phenotypes across multiple biobanks and health systems, followed by ancestry-specific and multi-ancestry meta-analyses, to provide adequately powered confirmation of these loci and refine credible gene and variant prioritisation.

To better understand genetic relationships between DED and systemic comorbidities, we performed a phenome-wide polygenic score association scan, followed by genome-wide genetic correlation analyses. Our results were largely consistent with known observational associations between DED and various systemic comorbidities, suggesting that shared genetic risk factors may partially underlie these relationships. However, neither polygenic score associations nor genetic correlations establish causal direction, and these findings should therefore be interpreted as evidence of shared genetic liability rather than causality. Notably, DED genetic risk exhibited multisystemic pleiotropy across neuropsychiatric, autoimmune, cardiometabolic, and chronic pain-related disease traits.

The overlap with chronic pain and psychiatric traits is particularly notable given increasing recognition that DED symptoms may not be fully explained by tear film dysfunction or ocular surface signs alone. Instead, persistent ocular discomfort may reflect interactions between peripheral ocular surface stressors and broader neurobiological susceptibility, including neuroimmune signalling, altered nociceptive processing, and peripheral or central sensitisation. Shared genetic liability with chronic pain and neuropsychiatric traits may therefore capture pathways related to immune activation, neuronal excitability, synaptic plasticity, stress responsivity, and pain amplification. This framework may help explain the frequent co-occurrence of DED symptoms with chronic pain and psychiatric comorbidity, particularly in patients whose symptoms are disproportionate to objective ocular surface findings.

Consistent with this interpretation, we jointly modelled several chronic pain traits using genomicSEM to derive a latent factor representing shared genetic architecture between DED and other chronic pain traits. The latent factor had a loading of 0.7 on DED, explained 51% of the genetic variance of DED, and itself exhibited multisystemic pleiotropy. Together, these findings suggest that DED may, in some patients, reflect broader multi-systemic health status rather than ocular surface pathology alone. A natural next step is to extend this genomicSEM framework to multi-factor models spanning neuropsychiatric, autoimmune, cardiometabolic, and pain-related traits, allowing more precise characterisation of the shared and domain-specific genetic pathways linking DED to systemic comorbidities.

Our findings help reframe the conversation regarding disease correlations, highlighting caution in imputing causation. Our results suggest that many traits are likely associated with DED through shared underlying aetiology, manifested as high genetic correlation. For instance, digital screen use has been associated with DED in numerous observational studies, with altered blinking dynamics and other mechanisms⁵⁵ proposed to contribute to this relationship. Yet our analysis suggests that associations between DED and screen use may be at least partially influenced by shared underlying factors. DED was associated with the polygenic score for overall screen use, but, when examining subtypes of screen use, we observed significant genetic correlations with TV viewing (r_G = 0.26; 95% CI: [0.19, 0.32]) and gaming (r_G = 0.15; 95% CI: [0.09, 0.22]), but not general computer use (r_G = 0.00; 95% CI: [−0.06, 0.06]). Notably, the chronic pain latent factor was even more genetically correlated with TV viewing (r_G = 0.39; 95% CI: [0.35, 0.44]) and gaming (r_G = 0.21; 95% CI: [0.17, 0.25]) than DED was, consistent with shared genetic or behavioural liabilities contributing to these associations. These findings do not exclude a causal effect of screen use on DED, but suggest that the observed associations may also partly reflect shared genetic liability, including genetic pleiotropy. An important future direction will be to use genetically informed causal inference methods, including Mendelian randomisation, to test the direction and magnitude of causal relationships between DED and related systemic traits. Bidirectional and multivariable MR approaches may be especially useful for distinguishing direct causal effects from shared pleiotropic pathways.

Our work has several limitations. First, our phenotyping algorithm was based on EMR diagnoses and prescriptions, which limits our ability to identify associations with specific DED symptoms and signs. Therefore, our findings should be interpreted as reflecting genetic liability to overall clinically recognised DED rather than to specific DED subtypes. An additional consideration is that our discovery GWAS used an EMR-derived definition of clinically recognised DED in MVP, whereas replication in GERA used related but not identical broad and narrow clinical definitions, and prior DED GWAS have included self-reported DED phenotypes. Self-reported symptoms and clinically diagnosed DED likely capture overlapping but non-identical aspects of a heterogeneous disease construct; therefore, modest differences in phenotype ascertainment across datasets may contribute to variation in genetic signal strength, effect-size estimates, replication performance, and cross-study heritability comparisons. Second, our study was performed in a population of military veterans receiving care at the VA, which may limit the generalisability of the results to other populations. This limitation is particularly relevant for DED, which is more prevalent in females, whereas the MVP cohort is predominantly male. As a result, our GWAS may be better powered to detect genetic risk factors that influence DED in male veterans, and the observed associations could be shaped by clinical, environmental, or comorbidity patterns that are more common in this population. Conversely, female-specific genetic effects may be underrepresented. However, our sex-stratified analyses provide some reassurance that the genetic architecture of DED is largely shared across sexes: we estimated a high genetic correlation between male and female DED which was not significantly different than 1 (r_G = 0.71; 95% CI: [0.41, 1.01]). Nevertheless, the confidence interval was wide, and these findings do not exclude the possibility of sex-specific loci or effect-size heterogeneity. Future work in larger cohorts with greater representation of women will be needed to validate these findings, improve power for sex-stratified analyses, and determine whether any DED risk loci have sex-specific effects. Finally, while our PGS scan and GWAS analyses were performed across all four major continental ancestry groups in MVP and demonstrated generally consistent results across ancestries, our heritability and genetic correlation analyses focused on EUR individuals. This was due to technical limitations of LDSC with respect to genetically admixed populations and a present lack of GWASs across many of the traits of interest in AFR and AMR populations. As these data and new methods become available over time, we plan to update this analysis to be more inclusive.

Strengths of this study include the development and validation of an EMR-based DED phenotyping algorithm, application of this algorithm in a large biobank across multiple genetically inferred ancestry groups, and integration of GWAS, heritability, phenome-wide PGS association, genetic correlation, and genomicSEM analyses. These complementary approaches enabled us to estimate common-variant heritability, identify DED-associated susceptibility regions, and characterise shared genetic liability between DED and a broader network of systemic, pain-related, autoimmune, psychiatric, and behavioural traits. Our results further motivate future analyses examining genetic correlations with specific signs and symptoms that comprise DED subtypes. Overall, our findings provide insights into the biological basis of DED and help unify years of observational associations of DED with systemic comorbidities and behavioural traits.

Contributors

Conceptualisation, B.R.G., S.K.I., N.S.P., and A.G.; Formal analysis, B.R.G., J.J.H., P.B.B., C.L.N., and C.W.H.; Validation, C.J. and H.C.; Writing—Original Draft, B.R.G., J.J.H., P.B.B., H.C., S.K.I., N.S.P., and A.G.; Writing—Review & Editing, B.R.G., J.J.H., P.B.B., C.W.H., C.L.N., C.C., M.F., P.B.G., W.C.W., H.C., T.B.B., S.K.I., N.S.P., and A.G.; Resources, S.P.; Funding acquisition, S.K.I., N.S.P., and A.G.; Supervision, T.B.B., S.K.I., N.S.P., and A.G. The VA Million Veteran Program contributed participant recruitment, biospecimen and genotype generation, electronic health record data infrastructure, and computational resources. B.R.G., J.J.H., P.B.B., C.W.H., C.L.N., and A.G. have accessed and verified the underlying data. All authors read and approved the final version of the manuscript.

Data sharing statement

Supplementary Tables in Excel format will be available upon publication through Zenodo (https://doi.org/10.5281/zenodo.22107554). Polygenic scores are available from the PGS Catalog (pgscatalog.org); URLs are provided in Supplementary Table S5. GWAS summary statistics used for genetic correlation and genomicSEM analyses are publicly available; URLs are provided in Supplementary Table S10. The full summary–level association data from the multi-ancestry meta-analysis and ancestry-specific MVP analyses will be available upon publication through GenHub (https://genhub.va.gov/public/sumstats). During the platform's initial deployment, users must authenticate through Login.gov or ID.me to meet VA information-security requirements; access does not require scientific review, project-specific approval, or a study-specific data-use agreement. The same summary statistics will also be available under controlled access through dbGaP under study accession phs001672, subject to dbGaP approval and applicable data-use agreements and restrictions.

Code sharing statement

The DED phenotype algorithm described in this article is available via open access through the Centralized Interactive Phenomics Resource (CIPHER) phenotype library at https://phenomics.va.ornl.gov. The exact prompt, PheWAS-based category labels, and Python script used with GPT-4 Turbo to classify PGS Catalog phenotype labels for visualisation are provided in the Supplementary Information. No additional custom analysis code was developed for the GWAS, PGS, LDSC, or genomicSEM workflows; these analyses were performed using the cited software, with software versions, model specifications, and key parameters reported in the Methods.

Declaration of interests

All authors declare that they have no competing interests.