Association between hot spring residency and dry eye disease: a crossover gene–environment interaction (GxE) study in Taiwan

Abstract

Background/aims

The advent of genetic biobanking has powered gene-environment interaction (GxE) studies in various disease contexts. Therefore, we aimed to discover novel GxE effects that address hot spring residency as a risk to inconspicuous disease association.

Methods

A complete genetic and demographic registry comprising 129,451 individuals was obtained from Taiwan Biobank (TWB). After geographical disease prevalence was analyzed to identify putative disease association with hot-spring residency, multivariable regression and logistic regression were rechecked to exclude socioeconomic confounders in geographical-disease association. Genome-wide association study (GWAS), gene ontology (GO), and protein-protein interaction (PPI) analysis identified predisposing genetic factors among hotspring-associated diseases. Lastly, a polygenic risk score (PRS) model was formulated to stratify environmental susceptibility in accord with their genetic predisposition.

Results

After socioeconomic covariate adjustment, prevalence of dry eye disease (DED) was significantly associated with hot spring distribution. Through single nucleotide polymorphisms (SNPs) discovery and subsequent PPI pathway aggregation, CDKL2 kinase pathways were significantly enriched in hot-spring specific DED functional SNPs. Notably, PRS predicted disease well in hot spring regions (AUC = 0.9168). Hot spring and discovered SNPs contributed to crossover GxE effect on DED (relative risk (RR)_G+E−= 0.99; RR_G-E+ = 0.35; RR_G+E+ = 2.04).

Conclusion

We identified hot-spring exposure as a modifiable risk in the PRS-predicted GxE context of DED.

Introduction

Gene-environment interaction (GxE) embodies a pathogenic context in which the impact of environmental exposures modifies our susceptibility to genetic mutations [1]. Although the GxE risk addition effects are reported to be potent and widely exist, GxE effects are often hard to discover and therefore overlooked from the scope of general pathophysiology. To this end, the advent of big data and the genetic biobanking era has enabled diverse GxE studies in the development of complex diseases [2]. Recent advances in translational exposomics have further highlighted the importance of integrating genomic and environmental data to personalize medicine and understand disease etiology more holistically [3].While environmental stimuli such as ultraviolet radiation, air pollutants (such as PM2.5, CO, and sulfates), and environmental toxins have been extensively studied in the context of GxE mechanisms [4], scant attention has been paid to the environmental effect of hot spring exposure—which allegedly can bring forth health benefits, hence we set a goal to understand the hot-spring related GxE interactions.

Throughout ancient cultures ranging across the empire of Rome, China, Japan, and the islanders of Austronesian, hot springs bathing was deemed to improve health, gain cosmetic benefits, and expand life. Recent scientific studies have gathered complementary evidence indicating that hot spring bathing could prevent diseases, alleviate musculoskeletal discomfort, modulate nerve and endocrine systems, and promote immunity against chronic diseases (such as whip syndrome, rheumatoid arthritis (RA), cancer, hemorrhoids, and skin diseases) [5, 6]. Moreover, Onsen therapy was certified by the Japanese Society for Complementary and Alternative Medicine (JCAM) [7], and was subsidized by health insurance in European countries (such as France, Germany, and Italy).

Hot spring is a geothermally heated water body that exerts health impacts through remodeling the local atmosphere and local hydrology. Bioactive gaseous substances such as H₂S generally range between 1.0 and 3.0 µg/m³ in urban areas, whereas the H₂S level spiked as high as 1500 µg/m³ in the residential vicinity of geothermal [8]. The exact effect of these bioactive geothermal-released substances was not fully understood, but independent pilot studies had reported a positive association between H₂S exposure and hospitalization events related to circulatory systems (Standardized Hospitalization Ratios (SHR) = 1.04; 95% CI = 1.01–1.07) [9]. Consistently, another geothermal epidemiology study in Italy also showed that increased H₂S exposure in men could result in higher cardiovascular mortality rates (odds ratio (OR) = 1.22, 90% CI = 1.03–1.44) [10]. Furthermore, a cohort study in Rotorua (New Zealand) showed that dwelling in geothermal emission areas was associated with increased malignancy rate and a 1.26 (95% CI = 1.14–1.38) standardized incidence ratio (SIR) for cataract [11]. While hot spring bathing was reported to bring folk-healing benefits to musculoskeletal diseases, hot spring residency was also linked to cardiovascular threats, therefore the health influence of hot spring exposure remains controversial to date.

To date, the relationship between geothermal exposure and ocular pathology remains under-explored. Hot springs can influence eye conditions through elevated temperature, humidity, and the emission of sulfide-containing aero-particles. Regarding these hot spring related exposome factors, it has been reported that low humidity can destabilize tear film quality in terms of tear break-up time (TBUT) and Schirmer test scores [12–16]. Interestingly, the often overlook geothermal gas emissions, such as nitrogen (N₂), carbon dioxide (CO₂), and hydrogen sulfide (H₂S), may impact tear film stability by affecting the lipid layer or acting as irritants [17]. While some studies link airborne pollutants to worsening dry eye symptoms [15, 18], a direct connection between geothermal-specific gases and tear film physiology is still uncharted.

Therefore, in this study, we set a goal to investigate the crossover effect between hot spring residency and genetic predisposition on ophthalmic diseases. To be specific, we performed a three-component study that included environmental association discovery, variant-disease association discovery, and GxE interaction analysis to delineate the environmental effects of hot spring exposure on ophthalmic diseases.

Materials and methods

Study design

This is a three-component study that includes environmental association discovery (demographic analysis, DA; and logistic regression model), variant-disease associationdiscovery (genome-wide association study, GWAS; protein-protein interaction analysis, PPI; and polygenic risk score, PRS), and GxE interaction analysis (crossover GxE effects and multivariate accumulative analysis).

Study population

129,451 individuals of Han Chinese ancestry were enrolled from general population by nation-wide TWB1 and TWB2 between 2016 and 2023 (https://www.twbiobank.org.tw/). At the time of entering the cohort, all participants were cancer-free and were obligated to complete a structured questionnaire that documented: sociodemographic characteristics, lifestyle, dietary habits, environmental exposures, and major medical histories [19]. Disease labels were self-reported, and control cases were designated by disease-free individuals above 60 years old.

Outcomes of interest

Primary outcomes of interest were set on hot-spring related ophthalmic diseases (cataracts; glaucoma; dry eye disease, DED; retinal detachment, RD; floaters; blindness; color vision deficiency, CVD; and age-related macular degeneration, AMD).

Exposure of interest

Exposure of interest included hot spring area residency and polygenetic underlying with adjustment of socioeconomic confounder. Hot-spring residential areas were townships (or districts) with hot spring outcrops documented by National Water Resource Department. Polygenetic underlying was assessed through variant-disease association (VDA).

Adjusted covariates for hot spring exposure and disease outcome

To find hot spring exposure related diseases, odds ratio (OR) and logistic regression were conducted to detect disease association with hot spring residency, and subsequent multivariate analysis was performed to solidify disease association under the consideration of covariate adjustments. Since hot springs are mainly distributed in townships with deviated sex ratios, imbalanced age composition, sparse medical institutes, and less income, socioeconomic and demographic covariates may confound with the relationship between hot spring exposure (E) and the disease outcome (D) (Supplementary Table S1 and Figure S1).

To specify the data source of the adjusted covariates: the demographic covariates (age, sex ratio, total resident counts) were recorded from the Directorate-General of Budget, Accounting and Statistics; total comprehensive income was collected from the released tolls of the Ministry of Finance (https://www.mof.gov.tw/); accessible medical resources were estimated by health insurance index (sum of health insurance benefits spent among all the hospitals in the township) and medical visit index (sum of annual medical visits among all the hospitals in the township) from the auditory releases of the Ministry of Health and Welfare (https://data.gov.tw/dataset/39280); altitude and relative humidity data was collected from automatic weather stations of Central Weather Administration (https://opendata.cwa.gov.tw/dataset/all/O-A0001-001); air pollution data were collected from air quality monitoring stations of Ministry of Environment (https://data.gov.tw/dataset/28202); UV radiation data was collected from ultraviolet rays monitoring stations of Ministry of Environment and Central Weather Administration (https://data.gov.tw/dataset/6076). Continuous variables that did not follow a normal distribution were provided with the mean ± standard deviation (SD) and quartile-based (min, Q1, Q2, Q3, and max) categorical conversion. The first category was pinned as reference, and then the effect of covariates was evaluated by univariable and multivariable logistic regression models.

Univariate logistic regression analyses were used to show the significant and independent role of the different variables in determining outcome (including floaters and DED) effects. Multivariate logistic regression analyses were used to control for confounding variables. All variables were then used for a multivariate logistic regression model with stepwise selection for constructing prediction models.

Since estimated coefficients in logistic regressions cannot be interpreted as a measure of the contribution of the effect, we have also calculated marginal effects. The marginal effect reflects the association between a small change in a variable and the alteration in probabilities across each outcome. These marginal effects were computed while keeping all other variables constant at the mean of the entire sample. Subsequently, multivariable logistic regression analyses were employed to investigate potential gene-environment interactions, utilizing the most suitable genetic model available. The predictive performance of the constructed predictive models was evaluated using sensitivity, specificity, and areas under the receiver operating characteristic (ROC) curves, the AUC value. All statistical analyses were performed using R version 4.1.1. The tests were 2-tailed, and p-value < 0.05 was taken as statistically significant.

Variant-disease association discovery

All SNPs discoveries were performed on PLINK (version 1.9 downloaded from http://zzz.bwh.harvard.edu/plink/). SNPs that does not meet following criteria may be removed from further analysis in the quality control process: (i) SNPs missing in more than 2% of participants (ii) individuals missing more than 5% SNP data (iii) SNPs with minor allele frequency (MAF) under 0.05 (iv) SNPs deviating from Hardy–Weinberg equilibrium (p-value < 10^− 6) (v) individuals with identity by descent (IBD) over 0.125.

GWAS identified disease-specific SNPs for populations living in townships with and without hot springs. SNPs were further screened by pairwise squared correlation (r²) less than 0.01 with linkage disequilibrium (LD) estimated in a window size of 4000 bp. Functional SNP genes were labeled as G_A for significant SNPs from all populations, G_N for significant SNPs from non-hot spring populations; G_H for significant SNPs from hot spring populations, and G_h indicated hot spring-specific functional SNP genes G_H that did not overlap in G_A and G_N.

To estimate an individual’s genetic risk of developing a disease, we constructed PRS models using the clumping and thresholding method in PLINK v1.9. Hyperparameters included clump-P, referring to the P value threshold for an SNP to be included as an index. We tested five clump-P values: P < 0.0100, P < 0.0075, P < 0.0050, P < 0.0010, and P < 0.0001. For each clump-P, SNPs within 25 Mb of the index SNP were considered for clumping, and SNPs with r² > 0.1 (we also tested r² = 0.1 and 0.01) with the index SNPs were removed. For SNPs with a negative logistic regression coefficient, we mathematically converted their minor alleles into major alleles, resulting in positive weight values for all variants [20]. The PRS was calculated as the sumof the logistic regression coefficients for all variants. We used threefold cross-validation for model building with each clump-P [21]. The TWB dataset was split into three equal, mutually exclusive subsets using R. Each subset served as the test set once, while the other four were used to train the model. This process was repeated three times. The model with the highest predictive performance, measured by the area under the ROC curve (AUC) using the R package “pROC,” was selected as the final model.

To demonstrate the genetic risk of disease predicted by the PRS, participants were divided into four groups based on their PRS. The (min, Q1) group included individuals with PRS values below the 25th percentile. The (Q1, Q2) group included those with PRS between the 25th and median percentiles. The same criteria were applied to the remaining two groups. Odds ratios for developing disease were calculated for each group, with the (min, Q1) group serving as the reference.

The functional PRS SNP genes were further analyzed on Enrichr [22–24] to attribute gene ontology (GO) and the ARCHS4 Kinases Coexp library was applied to identify co-expressed signaling pathways. PPI network analysis of G_h was further evaluated in the Search Tool for Retrieval of Interacting Genes/Proteins (STRING version 11.5 download from http://string-db.org/) [25] with a confidence level of 0.400 (medium) based on the find of: experiments, databases, co-expression, neighborhood, gene function, and co-occurrence. Then, the activated pathways in the selected PPI network were further calculated with the Reactome 2022 database.

GxE interaction analysis

Crossover GxE effects between environmental conditions and individual genetic variants were assessed by relative risk (RR) with binary variables: living in township with/without hot spring and high/low PRS score.

Visualized by nomogram, multivariate accumulative analysis based on logistic regression model estimated predictability of confounders (sex) and GxE effect. Decision curve analysis (DCA) evaluated clinical “net benefit” for nomogram prediction models in comparison to default strategies of treating all or no patients. Net benefit is calculated across a range of threshold probabilities, defined as the minimum disease risk at which further intervention would be warranted.

where w is the odds at the threshold probability [26].

Results

Finding hot spring associated diseases

This study included 129,451 participants with 4517 individuals dwelled in the townships with hot spring outcrops; their demographic characteristics were detailed as in Table 1.

Table 1.

Demographic information of the participants

	Hot spring residents (n = 4517)	Control (n = 124934)	Number (%)	p-value
Age (years)	48.56 ± 10.94	49.76 ± 10.96	129,451 (100.0)	< 0.001
Body mass index(kg/m2)	24.03 ± 3.40	24.00 ± 3.40	126,783 (97.9)	0.682
Water daily intake (ml)	1286.80 ± 633.15	1255.12 ± 685.91	21,735 (16.8)	0.07
Sex			129,451 (100.0)	0.392
Male (%)	1658 (36.7)	45,066 (36.1)
Female (%)	2859 (63.3)	79,868 (63.9)
Alcoholism			129,338 (99.9)	0.578
No or seldom (%)	4122 (91.3)	114,134 (91.4)
Abstain (%)	111 (2.5)	3268 (2.6)
Yes (%)	282 (6.2)	7421 (5.9)
Smoking experience			129,404 (100.0)	0.025
Yes (%)	3213 (71.2)	90,795 (72.7)
No (%)	1301 (28.8)	34,095 (27.3)
Meal times per day			129,447 (83.0)	< 0.001
1 (%)	3 (0.1)	302 (0.2)
2 (%)	139 (3.1)	4375 (3.5)
3 (%)	416 (9.2)	16,694 (13.4)
>3 (%)	6 (0.1)	113 (0.1)
Vegetarian			129,451 (100.0)	< 0.001
No (%)	506 (11.2)	19,415 (15.5)
Abstain (%)	30 (0.7)	1040 (0.8)
Yes (%)	28 (0.6)	1032 (0.8)
Physical activity			129,447 (83.0)	< 0.001
No (%)	2853 (63.2)	74,775 (59.9)
Yes (%)	1661 (36.8)	50,094 (40.1)
Physical activity kind			51,756 (40.0)	0.037
1 (%)	981 (56.5)	28,304 (56.5)
2 (%)	541 (32.6)	16,788 (33.5)
3 (%)	139 (8.4)	5003 (10.0)
Water origin			21,958 (17.0)	0.018
Tap water (%)	77 (1.7)	2716 (2.2)
Shallow well water (%)	0 (0.0)	13 (0.0)
Deep well water (%)	9 (0.2)	126 (0.1)
Mineral water (%)	35 (0.8)	1125 (0.9)
Filtered water (%)	390 (8.6)	15,934 (12.8)
Other (%)	49 (1.1)	1484 (1.2)

Continuous data are presented as means with standard deviation with Student’s t-tests (normally distributed data) or Wilcoxon tests (non-normally distributed data), and categorical data are presented as proportions with chi-squared tests or fisher exact test

Among the included disease spectrum, two diseases (floaters and DED) showed higher prevalence in hot-spring areas (Supplementary Figure S2). Subgroup analysis indicated that the classification of hot spring played an important role in disease risk (Fig. 1A). However, the coefficient of determination of these diseases were not significantly associated with hot spring distribution (Fig. 1B; R² = 2.51% in floaters and 1.54% in DED).

Fig. 1.

Discovering the environmental (E) effect on DED prevalence. A Association of hot spring subtype and ocular disease (DED and floaters). B The geospatial information system (GIS) map of hot spring-related ocular diseases. Each dot indicates a hot spring outcrop. C Logistic regression models evaluating socioeconomic (S) confounding effects. The dashed line indicated odds ratio (OR) = 1.0. D Logistic regression models evaluating other environmental confounding effects. The dashed line indicated odds ratio (OR) = 1.0. DED, dry eye disease

After univariate and multivariate logistic regression, DED (multivariable: OR = 0.78, 95% CI = 0.65–0.92; stepwise: OR = 0.81, 95% CI = 0.68–0.96) stood out as the remaining hot spring-associated disease (p-value < 0.05); whereas the hot spring association with floaters (p-value = 0.6807 for multivariable; p-value = 0.9935 for stepwise) was not statistically significant after adjusting for social and natural environmental factors (Fig. 1C and D, Supplementary Table S2). In these multivariate regression models, the average marginal effects (AMEs) of hot spring were statistically significant for DED (AME = − 0.23, 95% CI = − 0.40–0.07, p-value = 0.0060). Moreover, the incorporation of hot spring exposure information enhanced the stepwise logistic regression model performance in DED prediction (Supplementary Table S3). These statistic results indicated a robust association between DED risk modification and hot spring distribution.

Variant-disease association discovery

We applied GWAS (SNP list: Supplementary Table S4–S6) and Manhattan plot to visualize DED genetic predisposition. Interestingly, from the designated subgroup analysis, we noted that the top (p-value < 10^− 5) SNPs from non-hot spring subgroup were partial overlay with the parental all participants (for instance, rs6755439, rs7131682, rs78027155, rs10438549, AAK1 rs10439497, EPIC1 rs62227770, and so on) but were distinct to the hot spring subgroup findings. This SNP discovery discrepancy indicated a potential GxE effect (Fig. 2).

Fig. 2.

Manhattan plots of genome-wide association studies (GWASs) among “All”, “Non-hot Spring”, and “Hot Spring” gene sets. Blue lines represented p-value = 10^− 4 while red lines represented p-value = 10^− 5

We then calculated the predict AUC with the derived SNPs list at different significance (Fig. 3A). At p-value < 0.005, the DED AUC in hot spring subgroup was 0.92. However, in the non-hot spring subgroup, the model does not gain AUC increment over the enrolled SNPs.

Fig. 3.

Gene ontology. A Screening single nucleotide polymorphisms (SNPs) based on threshold of p-value < 0.005. AUC, Area Under Curve. B Protein‒protein interaction (PPI) network. C Comparing signaling pathways of co-expressed genes between gene set from G_A and G_h. The dashed line indicated p-value < 0.05. D Venn diagram of functional SNP genes from all, non-hot spring, and hot spring population. G_A: all population gene sets; G_N: non-hot spring population gene sets; G_H: hot spring population gene sets; and G_h: G_H that did not overlap with G_A and G_N. E Activated pathways of functional genes within PPI network of hot spring-specific gene set (G_h) based on Reactome 2022

After adjusting LD threshold (r² < 0.01) and genome-wide significance threshold (p-value < 0.005), the VDA results were aggregated and annotated as G_A, G_N, and G_H for “all”, “non-hot spring”, and “hot spring” subgroup respectively. 73 functional DED SNPs among G_A, G_N, and G_H were overlapped (Fig. 3D). After pathway discovery, there were 15 significant ARCHS4 human kinase pathways in the G_A of DED (adjusted p-value < 0.05) (Fig. 3C). However, CDKL2 kinase pathway was the only significant hit for DED G_h subgroup (overlapped ratio: 21/299 with GABARAPL2, C11ORF87, CXADR, DOCK3, ANKRD46, DDX25, PTPRO, ANK2, NALCN, LRP1B, MYO16, TRIM9, PCLO, CACNB4, DLG2, TRIM2, PSD3, TP53BP1, CTNNA2, SRGAP3, CDH18; OR = 2.79, adjusted p-value = 0.00698).

To investigate the putative molecular interaction in the derived G_h hits, we conducted Reactome pathway detection in the PPI after filtering and clustering G_h based on protein crosstalk in DED. Consisting of 158 genes with RYR2, CACNA1C, TGFβ1, and MAPK8 as four central hub genes, activated pathways in PPI network of DED were related to proposed DED-related pathomechanisms, such as O-linked glycosylation of mucins (R-HSA-913709), Interleukin (IL)-4 and IL-13 signaling (R-HSA-6785807), and interaction between L1 and ankyrins (R-HSA-445095)(Fig. 3B and E).

PRS models were constructed for different environmental conditions (all, non-hot spring, hot spring) based on the associated SNPs discovered from the TWB (G_A, G_N, and G_H). The mean PRS was higher among the cases compared to the controls across all models in both diseases but more significant in G_H (AUC = 0.9168 in Fig. 4A). AUC of PRS was higher than AUC of principal component analysis (AUC_PCA=0.5982 in Supplementary Table S7), indicating little population stratification effects on variant-disease association (VDA). Since the slope was 1.00 in the linear regression model of PRS SNP MAF between hot spring and non-hot spring regions (R² = 0.99, p-value < 0.001 in Fig. 4E), SNPs used in PRS model were independent of environmental conditions (E). However, PRS models with G_H could not predict disease risk well in populations without hot spring exposed (AUC = 0.5015 in Fig. 4D), indicating significant GxE relationship. The GxE relationship demonstrated a dose-response effect (Fig. 4B). Compared to the participants with PRS in min-Q1, participants with PRS in Q2-Q3 demonstrated 29.11 times odds of developing DED (95% CI = 10.36–81.76, p-value < 0.001) while the individuals with PRS in Q3-max demonstrated 390.81 times odds (95% CI = 130.93-1166.52, p-value < 0.001) (Fig. 4C). In threefold cross-validation, odds remained significant, 3.59 times in participants with PRS in Q2–Q3 demonstrated (95% CI = 1.92–6.73, p-value < 0.001) and 16.82 times in individuals with PRS in Q3-max (95% CI = 8.85–31.99, p-value < 0.001) (Supplementary Table S8).

Fig. 4.

Polygenic risk score (PRS) models evaluating gene-environment interactions (GxE). A Density distribution among PRS score. After adjusted for linkage disequilibrium (LD) clumping threshold (r² < 0.01) and genome-wide significance level threshold (p-value < 0.005), variant-disease association (VDA) results were depicted as G_A, G_N, and G_H for “all”, “non hot spring”, and “hot spring” gene sets. B Percentage of case and control in each PRS quartile. C Odds ratio (OR) of each PRS quartile. The dashed line indicated OR = 1.0. D Area Under Curve (AUC) of different combinations of gene set (G) and environmental condition (E). E Minor allele frequency (MAF) between hot spring and non-hot spring regions

GxE interaction analysis

While participants with different environmental conditions showed similar density profiles of G_HPRS (Fig. 5A), the disease ratio was significantly different with significant GxE interaction indicated by logistic regression model (p-value < 10^− 16), showing crossover GxE effect (Fig. 5C). Since disease ratio almost remained unchanged across G_HPRS, GxE effect was calculated based on risk in non-hot spring region as reference. RR tables between different environmental conditions and G_HPRS indicated hot spring as effect modifier (Fig. 5B). Therefore, we enrolled gender and GxE effect to construct DED predict nomograms (Fig. 5D). The DCA curve was performed to demonstrate high clinical net benefit between threshold probability of 20%-48% (Fig. 5E).

Fig. 5.

Net benefit of identifying the risk of gene-environment interactions (GxE). A Density profile of G_H polygenic risk score (PRS). B Multivariate logistic regression evaluating the GxE effect. C Relative risk (RR) with 95% confidence interval (CI) between different environmental conditions and G_HPRS. D Nomogram results estimating predictability of GxE. E Decision curve analysis (DCA) assessing net benefit of identifying GxE risk. PRS, polygenic risk score; Mo, mode; Md, median; G_H, gene set from “hot spring”; and DED, dry eye disease

Discussion

By employing multimodal analysis based on genetic data in TWB and DA in Opendata platform, we conducted environmental association discovery and variant-disease association discovery to reconstitute GxE effect on DED.

Hot spring and disease correlation

In the environmental association discovery, residency in hot spring areas was associated with a lower prevalence of DED (OR = 0.81). Resulted from increased osmolarity of the tear film and inflammation of the ocular surface [27], DED can be influenced by external factors such as desiccating environment and wind exposure [28]. In previous air pollutants studies, higher humidity levels were associated with improved corneal fluorescein staining and tear breakup time (TBUT) [15], while participants residing in areas with increased relative humidity were less likely to develop DED [29]. Given that hot spring water vapor may increase humidity levels in the immediate vicinity [30], it is possible that residing near hot springs were less likely to develop dry eyes. To the best of our knowledge, there are no studies publicly available to study the atmospheric influence on DED, future research avenue in the atmospheric (SO₂, CO₂ H₂S, N₂O) effect on DED occurrence.

Genetic effects in DED

In variant-disease association discovery, we demonstrated that co-expression of kinase pathways (CDKL2) and disease-related reactome pathways were significantly enriched in hot-spring specific functional SNPs.

Belonging to the family of CDC2-related serine/threonine protein kinases, cyclin-dependent kinase-like 2 (CDKL2) was identified as a target related to Sjögren’s syndrome (SjS) [31] and rheumatoid arthritis (RA) [32]. Moreover, DED is the most frequent ocular manifestation in patients with SjS (> 90%) and RA (46.7%) [33]. The co-expression of interleukins (IL-15RA, IL-7, IL-1RAPL1, and IL-11) also suggested the correlation between CDKL2 and SjS, RA, and DED, which are cytokine-mediated systemic inflammation [34]. Additionally, CDKL2 expression was detected in DED-involved ocular segments such as the lacrimal gland, cornea, and eyelid, according to the Human Eye Transcriptome Atlas [34]. Although a direct link between CDKL2 mutation and DED was not found in current research, we postulate that the effect of CDKL2 on DED is context-dependent and influenced by additional environmental cues.

Activated pathways in the PPI network of DED included O-linked glycosylation of mucins, IL-4 and IL-13 signaling, and interaction between L1 and ankyrins. Transmembrane mucins (such as MUC1, MUC4, and MUC16) [35], large glycoproteins with heavily glycosylated glycans, were revealed to be essential for maintaining ocular surface epithelium lubrication, wettability, and epithelial glycocalyx barrier. In response to inflammatory stimulation of DED, goblet cells and corneal/conjunctival epithelial cells express receptors for inflammatory cytokines such as IL-13 [36], inducing an important transcriptional factor in epithelial differentiation and goblet cell formation. Activation of transient receptor potential ankyrin 1 (TRPA1) resulted in calcium influx, tear reduction, increased eye blink, and forelimb eye wipe behavior [37]. These G_h-related pathways implied the underlying pathomechanisms of GxE in DED.

GxE interaction and clinical prospect

In this study, we reported that hot spring residency and underlying polygenetic conjointly modified the disease risk of DED, whereupon we gathered multi-angle evidence to strengthen the significance of GxE interaction. In addition to ophthalmic diseases, GxE interaction was also examined in other diseases (Supplementary Figure S2), such as orthopedic diseases (osteoporosis, arthritis, gout) and cardiopulmonary diseases (asthma, valvular heart disease (VHD), coronary artery disease, arrhythmia, cardiomyopathy, congenital heart disease, hyperlipidemia, hypertension, stroke, diabetes). However, only VHD showed similar GxE patterns with DED (Supplementary Figure S3).

Although it is generally accepted that both genetic and environmental factors contribute to the development of complex diseases, few GxE examples were replicated, biologically plausible, and methodologically sound interaction with proven clinical relevance and application in daily clinical routines. Therefore, the term “GxE interaction” was often used to express that several factors contribute to disease risk without excluding the possibility of complete independence. To be specific, biological GxE interaction is the joint effect of both genetic and environmental factors that act together in a direct physical or chemical reaction or in the same causal mechanism of disease development [38] while statistical GxE interaction is defined as “departure from additivity of effects on a specific outcome scale [39].” That is, the effect of one factor depends on the level of the other factor. The interaction could be defined as crossover effect or nonremovable interaction only when the interaction remained in any monotone transformation [40].

In this study, the interaction of hot spring residency and the level of G_HPRS well complied to the requirements of statistical GxE interaction. Interestingly, mediation analysis showed no significant differences in SNP MAF between hot spring and non-hot spring regions, while the G_HPRS density profiles varied across different environmental conditions (all, non-hot spring, and hot spring), which indicates that only under hot spring exposure could genetics variate the disease risk (Fig. 4). The presence of a crossover G × E effect was particularly striking, indicating that the relationship between G_HPRS and disease risk is influenced by environmental context (Fig. 5). In essence, the impact of genetic risk on disease likelihood varies depending on whether individuals are exposed to hot spring environments.

However, neither long-term residency effects nor diseases-specific effects have been thoroughly investigated with hot spring. Therefore, we struggled to interpret the biological GxE interaction via GO, co-expressed signaling pathways, and PPI network analysis. Overall, we proposed a novel GxE interaction in hot spring-associated diseases, whereby we suggest further studies are warranted to elucidate the benefits and harms of long-term hot spring residency in human health.

Study limitation

Despite the favorable results, there are several study limitations in this study. First, the acquisition of DED labels relied on self-reported questionnaires; the definitive diagnosis records and degree of disease severity were intrinsically absent from the biobank data. The hot spring exposure was determined by participant residencies, the exact exposure duration and intensity cannot be estimated and studied from existing accessible data. Due to the intrinsic data constraint, we were unable to retrieve individual-level exposure to the hot spring environment. Our initial aim was to provide a general estimation of the interaction between genetic background and the habit of residing in a hot spring-containing district. However, due to the administrative region-based classification of hot spring residence, individual proximity and exposure duration to the hot spring could not be accurately measured in our study. This limitation could be partially addressed through a more rigorously controlled study design, as demonstrated in a recent investigation. Specifically, Po-Chun Wang et al. proposed a feasible model that regulated both the temperature and exposure duration to hot water bathing among a cohort of Taiwanese patients with hypertension [41]. Nevertheless, achieving both a well-controlled behavioral experimental protocol and full transparency of genetic data remains challenging in real-world settings. A practical compromise may be to collect detailed residential address data, rather than relying solely on district-level information, to enhance the spatial resolution of this gene-environment (GxE) interaction study. Although multi-angle evidence supported the GxE interaction in hot spring-related DED, the concise pathomechanism cannot be inferred from statistical and epidemiological studies.

Together, this study served as a pilot study that highlights potential research avenues for GxE effects in hot spring-related diseases.

Author contributions

Conceptualization: Hsin-Yu Wu, Kao-Jung Chang; Data curation: Yi-Chen Lin, Yu-Tien Hsu, Yuan-Chih Wen, Pin-Hsuan Chiang, Yu-Hsiang Chen, He-Jhen Dai, Chia-Hsin Lu, Yi-Cheng Chen, Han-Ying Tsai. Formal analysis: Hsin-Yu Wu, Kao-Jung Chang, Wei Chiu, Ching-Yun Wang, Yi-Chen Lin, Pin-Hsuan Chiang, Yu-Hsiang Chen, Yi-Ping Yang; Funding acquisition: Yu-Chun Chen, Chih-Hung Hsu, Ai-Ru Hsieh, Shih-Hwa Chiou, Chih-Chien Hsu; Investigation: Hsin-Yu Wu, Kao-Jung Chang, Wei Chiu, Ching-Yun Wang, Yi-Ping Yang; Methodology: Hsin-Yu Wu, Kao-Jung Chang, Yi-Chen Lin, Pin-Hsuan Chiang, Yu-Hsiang Chen, Yi-Ping Yang; Project administration: Ai-Ru Hsieh, Chih-Chien Hsu; Resources: Yu-Chun Chen, Chih-Hung Hsu, Ai-Ru Hsieh, Shih-Hwa Chiou, Chih-Chien Hsu; Software: Yu-Chun Chen, Chih-Hung Hsu, Ai-Ru Hsieh, Shih-Hwa Chiou, Chih-Chien Hsu; Supervision: Yu-Chun Chen, Chih-Hung Hsu, Ai-Ru Hsieh, Shih-Hwa Chiou, Chih-Chien Hsu; Validation: Yu-Tien Hsu, Yuan-Chih Wen, He-Jhen Dai, Chia-Hsin Lu, Yi-Cheng Chen, Han-Ying Tsai; Visualization: Hsin-Yu Wu, Kao-Jung Chang; Writing—original draft preparation: Hsin-Yu Wu, Kao-Jung Chang, Wei Chiu, Ching-Yun Wan; Writing—review and editing: Hsin-Yu Wu, Kao-Jung Chang, Wei Chiu, Ching-Yun Wang, Yu-Tien Hsu, Yuan-Chih Wen, Ai-Ru Hsieh, Chih-Chien Hsu.

Funding

This work was supported by Medical scholarship foundation in memory of professor Albert Ly-Young Shen; Xu Jinde Memorial Foundation; Taiwan National Science and Technology Council [grant numbers NSTC 112-2314-B-032-001, 113-2321-B-A49-013 and 113-2314-B-305-001-MY2]; and Taipei Veterans General Hospital [grant number V113C-201, V112C-026 and 113VACS-007]. The funding sources had no involvement in study design; in the collection, analysis and interpretation of data; in the writing of the report; or in the decision to submit the article for publication.

Data availability

All data is published in this manuscript, the cited manuscripts, or the supplementary appendix. Data can be provided upon request to corresponding authors, and in agreement of terms. The datasets generated and/or analyzed during the current study are available in the Nation-wide Taiwan Biobank1 and Taiwan Biobank2 (https://www.twbiobank.org.tw/), Directorate-General of Budget, Accounting and Statistics (https://eng.dgbas.gov.tw) Ministry of Finance (https://www.mof.gov.tw/), Ministry of Health and Welfare (https://data.gov.tw/dataset/39280, https://data.gov.tw/dataset/39281, https://data.gov.tw/dataset/39282, https://data.gov.tw/dataset/39283) repository.

Declarations

Ethics approval and consent to participate

Ethical approvals were obtained from the Institutional Review Board (IRB) of the Taipei Veterans General Hospital with reference numbers 2023-01-006AC. All participants provided written informed consent. This study was conducted in compliance with the Helsinki Declaration. Patients or the public WERE NOT involved in the design, or conduct, or reporting, or dissemination plans of our research.

Competing interests

The authors declare no competing interests.