Predicting Risks of Dry Eye Disease Development Using a Genome-Wide Polygenic Risk Score Model

Abstract

Purpose: The purpose of this study was to conduct a large-scale genome-wide association study (GWAS) and construct a polygenic risk score (PRS) for risk stratification in patients with dry eye disease (DED) using the Taiwan Biobank (TWB) databases.

Methods: This retrospective case-control study involved 40,112 subjects of Han Chinese ancestry, sourced from the publicly available TWB. Cases were patients with DED (n = 14,185), and controls were individuals without DED (n = 25,927). The patients with DED were further divided into 8072 young (<60 years old) and 6113 old participants (≥60 years old). Using PLINK (version 1.9) software, quality control was carried out, followed by logistic regression analysis with adjustments for sex, age, body mass index, depression, and manic episodes as covariates. We also built PRS prediction models using the standard clumping and thresholding method and evaluated their performance (area under the curve [AUC]) through five-fold cross-validation.

Results: Eleven independent risk loci were identified for these patients with DED at the genome-wide significance levels, including DNAJB6, MAML3, LINC02267, DCHS1, SIRPB3P, HULC, MUC16, GAS2L3, and ZFPM2. Among these, MUC16 encodes mucin family protein. The PRS model incorporated 932 and 740 genetic loci for young and old populations, respectively. A higher PRS score indicated a greater DED risk, with the top 5% of PRS individuals having a 10-fold higher risk. After integrating these covariates into the PRS model, the area under the receiver operating curve (AUROC) increased from 0.509 and 0.537 to 0.600 and 0.648 for young and old populations, respectively, demonstrating the genetic-environmental interaction.

Conclusions: Our study prompts potential candidates for the mechanism of DED and paves the way for more personalized medication in the future.

Translational Relevance: Our study identified genes related to DED and constructed a PRS model to improve DED prediction.

Introduction

Dry eye disease (DED) is common among people of all ages and has attracted widespread attention over the past 30 years.¹ According to the latest definition by the Asia Dry Eye Society, DED is a multifactorial disease characterized by unstable tear film causing a variety of symptoms and/or visual impairment, potentially accompanied by ocular surface damage.² On the other hand, according to the definition in the Tear Film and Ocular Surface Society Dry Eye Workshop (TFOS DEWS) II report, DED is a multifactorial disease of the ocular surface characterized by a loss of homeostasis of the tear film and accompanied by ocular symptoms, in which tear film instability and hyperosmolarity, ocular surface inflammation and damage, and neurosensory abnormalities play etiological roles.¹ DED affects 5% to 50% of the adult population globally; in Taiwan, the prevalence of dry eye symptoms is estimated to be around 30% to 40%.³ Despite its high prevalence and impact on the quality of life, the lack of specific, sensitive, and objective criteria for diagnosis further complicates its research. Epidemiologic studies so far have identified numerous consistent non-modifiable risk factors associated with DED, including aging, female sex, Asian race, meibomian gland dysfunction (MGD), connective tissue diseases, and Sjögren's syndrome (SS). On the other hand, consistent modifiable risk factors associated with DED include androgen deficiency, computer use, contact lens wear, hormone replacement therapy, hematopoietic replacement stem cell transplantation, medications, and environmental factors, such as pollution, low humidity, and sick building syndrome.⁴ Additionally, procedures, nutrition, lifestyle challenges—such as physical factors and mental health—and even societal challenges impact ocular surface health.⁵ Nonetheless, little is known about the role of genetic factors and their relative importance in DED. 

Previously, genome-wide association studies (GWAS) have been conducted on two European populations, in the Netherlands and the United Kingdom.⁶ Single-nucleotide polymorphisms (SNPs) in the region rs4479326 (18q22.3) were identified to be associated with dry eyes (odds ratio [OR] = 0.70, 95% confidence interval [CI] = 0.66–0.75, P = 4.8 × 10⁻⁰⁸ for allele A). Small-scale investigations into non-SS dry eyes have uncovered potential roles of genes encoding pro-inflammatory cytokines, killer cell immunoglobulin-like receptors, and human leukocyte antigen-C.^7,8 Additionally, GWAS conducted on SS showed shared susceptibility loci in immune-related genes with other immune-mediated inflammatory diseases, such as systemic lupus erythematosus.⁹ Polymorphism variations in TNFα have been associated with both DED and SS.¹⁰ Animal studies have supported the role of inflammation in DED, exemplified by the autoimmune regulatory gene (AIRE) influencing lymphocyte subpopulations in the meibomian gland, phospholipid transfer protein (PLTP) affecting tear homeostasis and corneal permeability, and matrix metalloproteinase-9 (MMP-9) mediating corneal epithelial barrier function. Furthermore, the TNF receptor superfamily member 6 (Fas) gene and transforming growth factor-beta one (TGF-β1) are implicated in lymphocytes within the lacrimal gland, its secretory function, and autoimmune conditions such as SS. Despite these findings, the genetic landscape of DED remains elusive, and the validation of most genetic associations is still pending.¹¹ 

Using the Taiwan Biobank (TWB), a previous GWAS successfully identified two genetic risk loci for retinal detachment and developed a polygenic risk score (PRS) that effectively separated high-risk individuals from low-risk ones with genetic markers.¹² In this study, we conducted a large-scale GWAS for patients with dry eye using the TWB 1.0 and TWB 2.0 databases. The TWB database includes whole-genome genotype data of the Taiwanese population.¹³ The TWB (https://www.twbiobank.org.tw/new_web/) is a prospective cohort study of the Taiwanese population, who are mainly of Han Chinese ancestry, with genomic data divided into either hospital-based or community-based. The community-based biobank covering all the participants of this study includes repeated measurements of a wide range of phenotypes collected from more than 163,800 individuals (as of December 2022). The TWB recruits 30- to 70-year-old participants with no cancer history across 39 recruitment centers in Taiwan (the genetic ancestry of the subjects collected from the TWB is shown in Supplementary Fig. S1). In this study, 40,112 Han Chinese ancestry subjects were involved, and 16,900,296 loci were analyzed. Recognizing aging as a risk factor for DED and acknowledging the precedent set by numerous studies using 60 years as a cutoff, we examined the genetics in 2 distinct age groups: those younger than 60 years and those aged 60 years or older.^14–16 The 3,158,978 risk loci (<60 years old) and 3,149,919 risk loci (≥60 years old) for these patients with dry eye were identified at the genome-wide significance levels. Finally, we constructed a PRS for the risk stratification of patients with dry eye. The PRS was modeled and then verified with an independent cohort. We aimed to explore the genetic architecture of dry eye and generate a PRS model for risk stratification among individuals of Han Chinese ancestry. 

Methods

Study Population and Genome-Wide Association Study

This study population was exclusively obtained from the TWB, and similar methods have been adopted in a previous genetic study for retinal detachment conducting on the TWB.¹² After a participant was enrolled in the TWB, a structured questionnaire was collected covering demographics, lifestyle behaviors, environmental exposures, dietary habits, family history, and health-related information. In addition, the TWB developed two specialized SNP arrays for genotyping. The first array, TWBv1, was designed in 2011 using the Thermo Fisher Axiom Genome-Wide CHB Array as a base. It included approximately 650,000 markers on the GRCh37 coordinates, offering extensive coverage for common genetic variations suitable for GWASs. The second array, TWBv2, was designed in 2017 by Thermo Fisher Scientific and aimed to encompass not just GWAS markers but also functional variants, particularly rare coding risk alleles. Detailed genotyping and imputation procedures were described by Wei et al.¹⁷ 

As of December 2022, more than 163,800 participants were recruited. Among them, 114,600 participants have been genotyped with a custom TWB array. The inclusion criterion of cases was self-reported patients with DED. Control samples were individuals without DED. SNP quality control was conducted using PLINK (version 1.9) software to exclude SNP markers with missing call rate >5%, with minor allele frequency (MAF) <1%, or significantly deviated from the Hardy-Weinberg equilibrium (P < 1.0 × 10⁻⁶).¹⁸ Details of the quality control are shown in Supplementary Figure S2. After the quality control, a total of 16,900,296 variants from 40,112 participants (14,185 cases and 25,927 controls) were used in the subsequent analysis. Adjusting sex, age, body mass index (BMI), depression, and manic episodes as covariates, a logistic regression case-control analysis was performed. Institutional review board (IRB)/ethics committee approval was obtained. Informed consent was obtained from all subjects involved. 

It was known that age, sex, and SS participated in the pathogenesis of DED. To assess the differential genetic basis behind these factors, cases were further split into the younger group (<60 years old) and the older group (>60 years old), and genome-wide association analysis was performed on the two groups, respectively. For women who were more susceptible than men, analysis was also performed on female participants only. To assess the genetic structure of DED with SS, participants with self-reported dry eye and dry mouth symptoms were recruited as another subgroup for association analysis. 

Polygenic Risk Score Analyses

To build the PRS prediction models, the standard clumping and thresholding (C+T) method was adopted using PLINK (version 1.9).^18,19 The hyperparameters for this method were the cutoff of correlation r² and P value threshold. The parameter spaces for r² and P value threshold were 0.04, 0.2 and 7.5 × 10⁻⁴, 1 × 10⁻³, 2.5 × 10⁻³, 5 × 10⁻³, 7.5 × 10⁻³, and 1 × 10⁻², respectively. For each combination of r² and P value threshold, SNPs less than the P value threshold were retained. Secondly, SNPs with the lowest P value were chosen as the index SNP. Third, SNPs within the genetic window size of 250 kb around that index SNP were eliminated. The second and third steps were performed repeatedly until no SNPs were left. During PRS calculation, SNPs whose minor alleles carried protective effects on dry eye symptoms (with ORs less than 1) were converted into major alleles. Therefore, the new alternative allele was considered a risk allele (with an OR greater than 1), resulting in positive weight values for all variants. The predictive performances of PRS were demonstrated with the area under the receiver operating characteristic (ROC) curve (AUC) and were calculated using the R package pROC. Improvement in AUC between ROC curves was tested using Delong's method.²⁰ To prevent over-fitting, a five-fold cross validations was adopted. 

Results

Participant Characteristics

There were 40,112 participants from the TWB who were analyzed, including 14,185 self-reported patients with DED and 25,927 non-DED individuals. The 14,185 DED individuals were further divided into 8072 young (<60 years old) and 6113 old participants (≥60 years old), respectively. Participant characteristics are described in Table 1. Although women predominate among the DED participants (81.61%), the sex ratio is approximately 1 to 1 in non-DED participants. Furthermore, there are significantly higher proportions of depression and manic prevalence in DED individuals (7% and 1.4%) compared to non-DED individuals (3.41% and 0.54%, respectively). This is a reasonable prevalence in line with epidemiological studies.^21–23 

Table 1.

 

Basic Characteristics

 

Identification of DED-Related Loci

A GWAS was performed on 14,185 self-reported DED cases and the remaining 25,927 controls were used to identify DED risk loci. Overall, 17 SNPs exhibited adjusted P value < 1 × 10⁻⁵, implying their association with DED (Supplementary Table S1A, Supplementary Fig. S3). GWAS performed on young DED cases revealed 11 independent loci that showed an adjusted P value < 1 × 10⁻⁵ and could be matched with the genes ENSG00000286301 (rs2524181, adjusted P value = 7.27 × 10⁻⁸, and slope = 0.8761), DNAJB6 (rs9692250 adjusted P value = 8.93⁻⁸, and slope = 1.122), DCHS1, DTD1, LINC02267, MAML3, MICC, and ZFPM2 (Fig. 1A, Table 2, Supplementary Table S1B). In contrast, in old DED, 3 loci reached an adjusted P value < 1 × 10⁻⁵. These loci were located on the chromosomes 1, 2, and 13, with rs112017655 being the leading SNP (on 2q24.1, adjusted P value = 3.82 × 10⁻⁷; (see Fig. 1A, Table 2, Supplementary Table S1C). 

Figure 1.

 

(A) Manhattan plot illustrating the adjusted P values of SNPs tested for association with DED. (B) Manhattan plot illustrating the adjusted P values of SNPs tested for association with DED in female patients. (C) Manhattan plot illustrating the adjusted P values of SNPs tested for association with DED-carrying Sjogren-like syndromes. The upper part of each plot is an analysis of young DED cases, whereas the lower part is of old DED cases. The blue horizontal lines indicate the genomewide significance at adjusted P values = 1*10⁻⁵.

Table 2.

 

List of the Top SNPs Found to be Associated With Dry Eye Disease Via Genome Wide Association Study^a

 

Because women are more susceptible to DED due to different genetic architecture, we further performed a GWAS on female DED cases. Among the young female DED cases, 50 independent loci reached a P value < 1 × 10⁻⁷ (Fig. 1B, Supplementary Table S2A). The most significant locus was located on the chromosome 6p22.1 (rs2524181, slope = 0.8761, and P = 7.27 × 10⁻⁰⁸), and the second leading locus was located on the chromosome 7q36.3 (rs9692250 within the locus for the gene DNAJB6, slope = 1.122, and P = 8.93 × 10⁻⁰⁸). In contrast, in the analysis performed on old female DED cases, 17 independent loci showed suggestive evidence for association (with an adjusted P value < 1 × 10⁻⁶; Supplementary Table S2B). When comparing the differential genetic landscapes between young and old DED cases, 1 SNP showed suggestive association in both age groups, namely the rs62225456 SNP on chromosome 22q13.31 (P = 4.9 × 10⁻⁵). 

Identification of Risk Loci for DED With Symptoms of SS

SS is a systemic autoimmune syndrome characterized by dry eye and dry mouth (xerostomia) symptoms. To assess the genetic structure of DED with SS, participants with both dry eye and dry mouth symptoms were recruited as cases (n = 259), whereas participants without either symptom were regarded as controls (n = 415). Six independent loci reached adjusted P < 10⁻⁵ in the young DED group, leading by rs13111065 on chromosome 4 (adjusted P value = 8.10 × 10⁻⁷; Fig. 1C, Supplementary Table S3A). Five independent loci exhibited adjusted P < 10⁻⁵ in old DED group (see Supplementary Table S3B). There were no overlapping loci with shared significance in both the young DED group and the old DED group. 

The Polygenic Risk Score and DED Risk Stratification

To stratify the risks of DED development, we constructed a PRS model for DED in old and young populations based on associative SNPs identified in GWAS. Tuning parameters included adjusted genome-wide significance level (adjusted P value) and linkage disequilibrium clumping threshold (r²). The models’ performance based on various combinations of parameters is shown in Table 3. Mean PRS across all models were higher in DED cases than in non-DED participants, indicating the role of genetics in DED development. Based on model performances and clinical usefulness, the final model was built with P < 2.5 × 10⁻³ and r² < 0.04. For young DED, 932 SNPs were included; for old DED, 740 SNPs were utilized. 

Table 3.

 

Polygenic Risk Models for Stratifying Dry Eye Disease Patients From Controls Based on Different Tuning Parameters

 

The PRS effectively distinguishes the risks of DED development (Fig. 2). For the young population, the top PRS quantile (Q3-max) had a 17.5-fold (OR = 17.5 [15.64, 19.67]) risk of acquiring DED than the lowest PRS quantile (min-Q1). The second (Q2-Q3) and the third highest PRS quantiles (Q1-Q2) exhibited 6.00- and 2.23-fold risks compared to the lowest PRS quantile, suggesting the dose-effect relationship between the PRS value and the risks of acquiring DED. A similar result was also found in the old population. The ORs of DED development were 17.85, 5.63, and 2.57-fold in the highest, second highest, and third highest PRS quantiles, respectively, compared to the lowest PRS quantile. 

Figure 2.

 

Distribution of cases and controls according to the PRS. (A) Distribution of the PRS in young DED cases and controls. (B) Distribution of young DED cases and controls based on PRS quantiles. (C) Odds ratio for acquiring DED based on PRS quantiles. (A–C) were for younger cases (age <60 years), whereas (D–F) were for older cases (≥60 years).

Furthermore, our PRS model shows statistically significant value in identifying high-risk individuals. For the young population, individuals in the top 5% of the PRS had a 10.05-fold risk, the top 10% exhibited a 7.81-fold risk, and the top 25% had a 5.83-fold greater risk of DED than the remaining individuals. For the older population, individuals in the top 5% exhibited a 10.12-fold risk, the top 10% had a 7.62-fold risk, and the top 25% had a 6.09-fold greater risk of DED than the remaining individuals. 

The ROC curve was used to assess the performance of our PRS models. In the young population, the area under the ROC curve (AUROC) for the PRS model alone was 0.537 (blue curve in Fig. 3A). When DED-related covariates, namely age, sex, BMI, depression, and manic episodes, were added to the models, the AUROC increased to 0.648 (green curve). For the old population (Fig. 3B), the AUROC for PRS alone was 0.509 (blue curve), and the value increased to 0.600 when covariates were included in the model (green curve). Furthermore, the AUROC for the final testing model that included PRS and covariates, were 0.6484 and 0.5999 for young and old DED, respectively. 

Figure 3.

 

Receiver operating characteristic (ROC) curve of the polygenic risk scores for predicting DED for (A) young cases (<60 years old) and (B) old cases (≥60 years old). The orange curve represents the predicting ability using clinical covariates, whereas the blue curves use PRS, and the green curves use both clinical covariates and PRS. The area under curve (AUC) is illustrated in the figure.

Discussion

The TWB is a nationwide research database that integrates the demographic, phenotypic, and genetic data of over 172,000 Taiwanese. Using the biobank, we sieved out 14,185 self-reported cases and 25,927 controls to investigate the genetics of DED. All controls were 60 years old and over. The cases were split into young DED cases and old DED cases, because the genes that impact DED may vary among young and old individuals. GWAS was performed on the young and old cases separately. 

Risk Loci for Dry Eye Disease

Among the young DED cases, we identified 11 independent risk loci, including DNAJB6, MAML3, LINC02267, DCHS1, SIRPB3P, HULC, MUC16, GAS2L3, and ZFPM2. MUC16 encodes a protein that is a member of the mucin family. Mucins play many important roles on the ocular surface as a protective mucous barrier, including the maintenance of lacrimal fluid, lubrication of the ocular surface to facilitate blinking of the eye, formation of a smooth spherical surface for good vision, and trapping and removing pathogens and debris. At least four subtypes of secreted mucins (MUC5AC, MUC7, MUC2, and MUC19) and four types of membrane-associated mucins (MUC1, MUC4, MUC16, and MUC20) are expressed on the ocular surface on the mRNA and/or protein levels.²⁴ A previous study showed that postmenopausal women with a history of DED displayed significantly increased cellular MUC16 protein levels (P < 0.001). This may indicate a compensatory effect to the irritation and inflammation caused by the disease.²⁵ Mucin plays an important role in DED, as it makes up tear film and tear film instability is the main mechanism behind DED. Currently, tear film-oriented therapy has become the core concept of DED treatment. Mucin secretagogues such as diquafosol sodium and rebamipide eye drops had been a well-known TFOT for patients with DED with mucin deficiency.² 

Several other genes harboring the higher risk-associated SNPs can be potentially associated with the pathogenesis of DED. Among them, ADAMTS17 was found to be related to the Weill-Marchesani-like syndrome, which is a rare genetic disorder that affects the musculoskeletal system, the eyes, and the cardiovascular system. It is characterized by multiple eye disorders including lenticular myopia, ectopia lentis, glaucoma, and spherophakia, as well as short stature. Consistently with the fact that rheumatoid arthritis (RA) is a risk factor for dry eye, ADAMTS17 was found to be hypomethylated in B lymphocytes of patients with RA.²⁶ 

The product of PTPN2 gene, also known as T cell protein tyrosine phosphatase (TCPTP), dephosphorylates the insulin receptor, EGFR, Src family kinases, JAK, and STAT. PTPN2 regulates the interactions between innate and adaptive immune cells and intestinal epithelial cells.²⁷ PTPN2 mutation leads to macrophage polarization to M1 phenotype, which shows better pro-inflammatory ability than M2 macrophages. Hyperosmotic stress also promotes M1 polarization, and a greater proportion of activated M1 macrophages were observed in the conjunctiva tissue in the DED mouse model.²⁸ Anti-IL-20 treatment, on the other hand, suppresses hyperosmolarity-induced pro-inflammation and contains M1 polarization.²⁹ Furthermore, PTPN2 mutations were shown to limit the formation of autophagosomes and increase the permeability of the intestinal epithelial barrier. PTPN2 is also known to be related to lymphoid malignancies and neuronal development.^30,31 

Next, the Ubiquitin C-Terminal Hydrolase L1 (UCHL1) gene encodes a peptidase that is specifically expressed in neurons and neuroendocrine cells. So far, mutations in UCHL1 have been linked to Parkinson's disease and autosomal recessive spastic paraplegia and optic atrophy.³² T cells expressing UCHL1 were also found in the malignant lymphoma of a patient with SS.³³ 

OR11A1 and OR5V1 are the genes encoding the olfactory receptors. Olfactory receptors are largely intertwined with many neurotransmitters and hormone receptors, responsible for the recognition and transduction of odorant signals. The olfactory receptor gene family is notably the largest in the genome. Olfactory dysfunction is commonly found in patients with primary SS and there is clinical evidence that they seem to be correlated with dry eye parameters.³⁴ 

SHLD2, also known as FAM35A, plays an important role in the repair of DNA double-stranded breaks. In previous population-based studies, mutations in SHLD2 have been found to be associated with hyperuricemia and gout.²⁸ Gout is a less known risk factor of DED.^35,36 

Finally, we highlight the potential role of TRANK1 in DED that remains to be explored. TRANK1 is a bipolar disorder susceptibility gene.³⁷ Interestingly, bipolar disorder has been associated with DED,^37,38 including a study conducted among the Taiwanese population (OR  =  1.9).³⁹ 

ZFPM2, CACNA2D3, DNAJB6, and TRANK1 are other risk loci, but there have been no reported correlation between these genes and DED (Supplementary Table S4). 

Notably, the previous studies showed that gender-related sociocultural factors are at play; women are more likely to seek medical care when experiencing DED-related discomfort.⁴⁰ Sex-specific biological differences are also implicated, whereby sex hormones may influence DED.⁴¹ We therefore focused on the genetics of DED in the female population. Among the risk loci associated with the young patients with DED, risk loci at ENSG00000286301, DCHS1, DNAJB6, MAML3, MICC, and ZFPM2 were the most significantly associated with young female patients with DED; among risk loci associated with the old DED cases, only ENSG00000286679 was significantly associated with the old female cases. All the genes identified in the female-specific GWAS are a subset of the genes identified in the overall GWAS, suggesting the differential role of genes in the DED pathogenesis among women. 

Risk Loci for Dry Eye Disease With Dry Mouth

Dry eye is a common symptom of several conditions, including SS which is characterized by both dry eye and dry mouth. Therefore, we investigated the genetics of individuals with both dry eye and dry mouth. The identified genes were C1QL1, DCHS2, EIF4E3, KCNB2, and THSD7A in young patients. All genes identified in this population were different from the previously identified genes, suggesting that this is a distinct subpopulation. 

The complement component 1, Q subcomponent-like protein (C1QL1), in contrast to C1q, has no functions in the complement cascade. However, it is expressed at significantly higher mRNA and protein levels in Th17 cells than in the CD4+ T cell subtypes. The proteins they encode for are well associated with the inflammatory function of Th17 cells.⁴² Given that T cells function in mediating DED,⁴³ the role of C1QL1 in DED is worth exploring in future studies. 

The dachsous cadherin-related 2 (DCHS2) gene encodes a large protein that likely functions in cell adhesion. E-cadherin is known to be important for meibomian gland function and the serum levels of soluble E-cadherin were elevated in primary SS.^44,45 However, the role of DCHS2 in SS or DED remains to be investigated. 

Ethnic variations have been observed in the natural progression of DED. In comparison to the Caucasian population, middle-aged and older individuals of Asian descent exhibit a shorter noninvasive tear-film break-up time, higher tear film osmolarity, increased fluorescein corneal and conjunctival staining, elevated meibography grade, and a greater expressed meibum grade. Notably, the tear meniscus height was found to be similar in both ethnicities, suggesting that aqueous tear deficiency is less likely to contribute to the DED in the Asian population than evaporative mechanisms, specifically those associated with MGD.⁴⁶ The genes identified in this GWAS, some of which may play a role in Meibomian gland function, inflammation mediation, tear film composition, or neurotransmitter regulation, could offer insights for future investigations correlating differential DED etiologies among various ethnicities. 

In addition to exploring the molecular mechanisms of DED, we also constructed a PRS score in hopes of creating a clinical tool that can predict the risk of DED development. Nine hundred thirty-two and 740 genetic loci were incorporated into the PRS model for the young and old populations, respectively. The higher the PRS score, the higher the risk of acquiring DED. Individuals within the top 5% of PRS exhibited around 10 times higher risk of developing DED. Therefore, in the era of genetics, early identification of DED is feasible using this PRS model. Early education and prompt medical interventions can prevent irreversible corneal injury. Furthermore, DED is a multifactorial disease influenced by age, sex, and BMI, and all these factors should be accounted for in the PRS model. After integrating these covariates into the PRS model, the AUROC increased from 0.509 and 0.537 to 0.600 and 0.648 for young and old populations, respectively, demonstrating the interplay between genetic and environmental factors. Overall, the major advantage of this study is the big size and relative racial homogeneity of the cohort. This study evaluated various DED-related factors that are available in the database and identified the SNPs for DED and female patients with DED. This study encountered several limitations. First, identifying DED cases relied on self-reported status rather than utilizing more objective measures, such as the Schirmer test, fluorescein dye, or tear break-up time. The absence of ophthalmic investigations and diagnoses by an ophthalmologist during participant recruitment could potentially lead to both missed and overestimated DED cases. Additionally, discrepancies between observed DED signs and reported symptoms were observed.⁴⁷ Nevertheless, given the high accessibility of ophthalmologic clinics in Taiwan facilitated by the national health insurance program and the predominant reliance on dry eye symptoms for DED diagnosis, the utilization of self-reported disease status is deemed reasonable for the scope of this study and is unlikely to introduce significant bias. It is acknowledged that the incorporation of more objective and quantifiable examinations, if feasible, could enhance the robustness of the study outcomes. Second, DED is often loosely defined in the clinic. The poor classification of etiologies of DED, such as allergy and SS, is limited by data accessibility and its heterogenous nature. Clinically, it is a common practice not to distinguish among the lipid-deficiency, mucin deficiency, and aqueous-deficiency types of DED. Furthermore, the unavailability of data on anti-Ro/SSA, anti-La/SSB antibodies, and other elements in the SS classification criteria in our databank constrained our exploration of the genetic aspects of SS within the context of DED. 

Overall, our study was based on a total of 40,112 samples from the Taiwanese population, with 14,185 DED individuals and 25,927 healthy controls. The GWAS disclosed 11 risk loci associated with DED and 5 risk loci associated with both dry eye and dry mouth, which are different from the Western population. Otherwise, our PRS prediction model provides a chance to improve case detection. Validation studies with other populations are required to confirm our findings. In addition, functional analyses are needed to clarify the contribution of the identified loci to the development of DED. Nevertheless, our study prompts potential candidates for the mechanism of DED and paves the way for more personalized medication in the future. 

Acknowledgments

The authors thank the Big Data Center (BDC) of Taipei Veterans General Hospital (VGHTPE), and the Department of Statistics, Tamkang University for technological support. 

Supported by the grants of the National Science and Technology Council (NSTC 112-2321-B-A49-007), Taipei Veterans General Hospital (V113C-201), VGHUST Joint Research Program (VGHUST112-G1-5-1), the Veterans Affairs Council (110VACS-007), Center for Intelligent Drug System and Smart Bio-device (IDS2B) from the Featured Areas Research Center Program within the framework of the Higher Education Sprout by the Ministry of Education, Taiwan. 

Disclosure: C.-C. Hsu, None; H.-K. Chuang, None; Y.-J. Hsiao, None; P.-H. Chiang, None; S.-W. Chen, None; W.-T. Luo, None; Y.-P. Yang, None; P.-H. Tsai, None; S.-J. Chen, None; A.-R. Hsieh, None; S.-H. Chiou, None