Active Cigarette Smoking Is Associated With an Exacerbation of Genetic Susceptibility to Diabetes

Diabetes influences the lives of hundreds of millions of people around the world, and its prevalence is gradually increasing (1). Genetic factors play an important role in the development of diabetes. Fasting glucose (FG) and glycated hemoglobin (HbA_1c) are commonly used for the diagnosis of diabetes. An FG level >126 mg/dL (2) and an HbA_1c level >6.5% (48 mmol/mol) (3) have been proposed as diagnostic indicators for diabetes. The heritability levels of FG and HbA_1c were estimated to be 51% and 62%, respectively (4). The remaining variability in FG and HbA_1c could be explained by lifestyle factors.

Smoking has been found to be associated with a higher risk of insulin resistance and diabetes (5). Nicotine is thought to be responsible for the association between cigarette smoking and the development of diabetes (6). However, smoking has not been uniformly regarded as a risk factor of diabetes (7). Moreover, it is important to uncover whether smoking may modify the genetic risk of diabetes.

To know whether the genetic risk of diabetes may vary with cigarette smoking, we here investigated gene-smoking interactions (G×S) on FG, HbA_1c, and the dichotomous diabetes status, respectively. Our Taiwan Biobank (TWB) data included a discovery cohort (TWB1) and a replication cohort (TWB2). Genetic risk score (GRS) of each TWB2 subject was calculated with weights retrieved from TWB1 analyses. We tested whether the GRS effects could be modified by active smoking status, the number of pack-years, years as a smoker, packs smoked per day, or hours as a passive smoker per week. This study aims to uncover whether smoking can modulate the genetic predisposition to diabetes and which smoking measurement is the most critical effect modifier.

The TWB collects genomic and lifestyle information from Taiwan residents aged 30–70 years (8). To participate in the TWB, community-based volunteers signed informed consent, took physical examinations, and provided blood and urine samples. Their lifestyle factors were further collected through a face-to-face interview with TWB researchers. TWB received ethical approval from the Ethics and Governance Council of Taiwan Biobank, Taiwan, and also from the Institutional Review Board on Biomedical Science Research, Academia Sinica, Taiwan. Our study was approved by the Research Ethics Committee of National Taiwan University Hospital (NTUH-REC no. 201805050RINB).

This study included 27,737 and 67,512 subjects who were whole-genome genotyped by the TWB1 and TWB2 genotyping arrays, respectively, until February 2020. To explore cryptic relatedness, we used PLINK 1.9 (9) to estimate cryptic relatedness, i.e., PI-HAT = Probability (IBD = 2) + 0.5 × Probability (IBD = 1), where IBD is the genome-wide identity by descent (IBD) sharing coefficients between any two TWB individuals. Similar to many genetic studies (10,11), we also excluded relatives by removing one subject from every pair with PI-HAT ≥0.2. A total of 25,460 unrelated TWB1 subjects and 58,774 unrelated TWB2 subjects remained after this quality control process.

Most TWB subjects were of Han Chinese ancestry (8). Released in April 2013, the TWB1 genotyping array was designed for Taiwan’s Han Chinese and run on the Axiom Genome-Wide Array Plate System (Affymetrix, Santa Clara, CA). On the basis of user experience in TWB1 and next-generation sequencing of ∼1,000 TWB individuals, the TWB2 genotyping array was released in August 2018 to further cover specific single nucleotide polymorphisms (SNPs) in Taiwan population. There were 632,172 and 648,611 autosomal SNPs that were genotyped in TWB1 and TWB2 arrays, respectively. In TWB1, we removed 27,628 SNPs with genotyping rates <95%, and 6,900 SNPs with Hardy-Weinberg test P values < (12). In TWB2, we removed 27,859 SNPs with genotyping rates <95%, and 14,656 SNPs with Hardy-Weinberg test P values < (12). Finally, 597,644 TWB1 SNPs and 606,096 TWB2 SNPs were kept in our analysis and were used to construct ancestry principal components (PCs). Between these two arrays, 92,052 SNPs were overlapped.

We imputed the genotypes of autosomal SNPs using the Michigan Imputation Server (https://imputationserver.sph.umich.edu/index.html), with the reference panel based on the East Asian population from the 1000 Genomes Phase 3 v5. After removing SNPs with Hardy-Weinberg test P values < (12) and SNPs with low imputation information score (R² <0.8), 7,433,014 and 6,347,468 SNPs remained in TWB1 and TWB2, respectively.

We analyzed two continuous traits related to the diagnosis of diabetes: FG and HbA_1c. After a minimum 6-h fast (no calorie intake for at least 6 h), serum glucose was measured using a Hitachi LST008 analyzer (Hitachi High-Technologies, Tokyo, Japan), whereas HbA_1c was measured with the Trinity Biotech Premier Hb9210 analyzer (Bray, Ireland/Kansas City, MO). As in the criterion used worldwide (2,3), the Ministry of Health and Welfare in Taiwan defined an FG level higher than 126 mg/dL or an HbA_1c level higher than 6.5% (48 mmol/mol) as an indicator of diabetes.

Moreover, diabetes status was also analyzed as a dichotomous trait, where subjects with diabetes included those with physician-diagnosed diabetes, or those having FG >126 mg/dL or HbA_1c >6.5% (48 mmol/mol) according to the TWB test results.

In TWB, an active smoker was defined as a subject who had actively smoked for at least 6 months and had not quit smoking at the time his or her FG and HbA_1c were measured. Moreover, TWB researchers also asked every active smoker how many packs of cigarettes were smoked per day and how many years they had been an active smoker. The number of pack-years was then calculated as multiplying the number of packs smoked per day by the number of years as a smoker. Furthermore, TWB researchers asked each individual the number of hours they had as a passive smoker per week. In total, five smoking measurements will be investigated in the following analyses.

Sex, age, BMI, alcohol drinking status, regular exercise, and educational attainment were regarded as covariates and were adjusted in all of our regression analyses. Drinking was defined as a subject having a weekly intake of more than 150 mL of alcohol for at least 6 months and having not stopped drinking at the time his or her FG and HbA_1c were assessed. Regular exercise was defined as engaging in 30 min of exercise three times a week. Exercise indicates leisure-time activities, such as jogging, mountain climbing, yoga, and so on (13).

Educational attainment was recorded through a face-to-face interview with TWB researchers. It was represented by a number ranging from 1 to 7, where 1 indicates illiterate; 2 means no formal education but literate; 3 denotes primary school graduate; 4 represents junior high school graduate; 5 indicates senior high school graduate; 6 means college graduate; and 7 denotes master’s or higher degree. Similar to a recent study for diabetes (14), we also considered educational attainment as an covariate in all analyses.

Initially, to test whether smoking was significantly associated with FG, HbA_1c, and diabetes, we regressed each trait according to the following model:

where Y is natural log transformed FG (or HbA_1c) with an identity link ⁠, or the diabetes status with a logit link ⁠. Smoking is coded as 1 for active smokers and 0 otherwise.

Sex, age, BMI, educational attainment, and ancestry PCs are commonly adjusted in models for FG, HbA_1c, and diabetes (14,15). Regular physical exercise has been shown to reduce the risk of insulin resistance, metabolic syndrome, and diabetes (16). Effects of cigarette smoking and alcohol consumption on diabetes are usually investigated (5–7,17). Therefore, a total of 16 covariates were adjusted in model 1, including sex, age (in years), BMI, drinking status (yes vs. no), regular exercise (yes vs. no), educational attainment (a value ranging from 1 to 7), and the first 10 ancestry PCs.

Because FG and HbA_1c were skewed to the right, both traits were natural log transformed to improve the R² of model 1. Such a transformation was commonly used in analyses for FG and HbA_1c (18,19). On the basis of model 1, we later replaced active smoking status with the four continuous smoking measurements.

To build a trait-specific GRS, we first identified FG-associated SNPs, HbA_1c-associated SNPs, and diabetes-associated SNPs based on the 25,460 unrelated TWB1 subjects. Each trait (denoted by Y) was regressed on every SNP while adjusting for the 16 covariates, as follows:

⁠, the P value of testing from model 2, calibrates the significance of marginal association of SNP j with Y. Because of ∼8 million SNPs in TWB1, SNP j is significantly associated with a trait if ⁠. We calculated GRS for each of the 58,774 unrelated TWB2 subjects by

where was estimated from model 2, and SNP_j is the number of minor alleles at the jth SNP (0, 1, or 2) for the TWB2 individuals. The indicator function is 1 if and 0 otherwise.

Because is composed of SNPs that are more associated with FG, HbA_1c, and diabetes (satisfying ⁠), this is the so-called marginal-association filtering in gene-environment interaction analyses (20–23). To avoid collinearity in a GRS, we used the default setting of the clumping procedure in PLINK 1.9 (9). To be specific, if there were multiple variants with within a 250-kb region, we kept the most significant one (so-called index variant) and removed others in linkage disequilibrium with the index variant (r² > 0.5). To quantify how many SDs) a is from the mean, we performed the z score transformation on and then obtained the GRS. By fitting model 2 for FG, HbA_1c, and diabetes separately, we obtained the GRS for FG, HbA_1c, and diabetes, respectively.

SNP_j is the number of minor alleles at the jth SNP (0, 1, or 2). A positive indicates that the minor allele is trait-increasing, and a subject with more copies of the minor allele (more trait-increasing alleles) will receive an increment in his or her GRS. In contrast, a negative represents that the minor allele is trait-decreasing, and a subject with more copies of the minor allele (more trait-decreasing alleles) will obtain a decrement in his or her GRS. Finally, a higher GRS is associated with a larger degree of FG, HbA_1c, and diabetes.

In addition to the TWB1-derived GRS previously mentioned, we also calculated a European-derived GRS (EuGRS) based on 38 diabetes-associated SNPs that were previously identified from genome-wide association studies (GWASs) (24–27). Most of these 38 SNPs were identified at the commonly used genome-wide significance level ⁠. Individuals of those GWASs (24–27) were overwhelmingly of European descent. The European-derived GRS was calculated as ⁠, where the weights (⁠⁠) were the effect sizes reported by previous GWASs (24–27) and summarized by Said et al. (14).

To present an analysis parallel to GRS-smoking interaction analysis (where TWB1 was used to find trait-associated SNPs and TWB2 was used to test for interactions), EuGRS-smoking interaction analysis was also performed on the TWB2 cohort with 58,774 subjects. There is one more reason why we did not include the TWB1 cohort into the EuGRS-smoking interaction analysis. While 32 out of the 38 SNPs were genotyped by the TWB2 array, only 27 were genotyped by the TWB1 array. After imputation, still two SNPs in TWB1 failed to achieve the criterion of imputation information score (i.e., their R² < 0.8). Therefore, we did not compute EuGRS for the 25,460 TWB1 subjects. To sum up, both GRS-smoking interaction and EuGRS-smoking interaction were evaluated in the TWB2 cohort with 58,774 subjects.

We then used the GRS to test for the presence of ⁠. The following model was considered for each trait:

where Y is natural log transformed FG and HbA_1c or the diabetes status, and GRS is the standardized GRS z score of the TWB2 subjects. Covariates adjusted for FG, HbA_1c, and diabetes have been described under model 1. To further control confounding (28), GRS × Covariate_v and Smoking × Covariate_v (⁠⁠) were also included in model 4. To interpret the main effects in model 4, we centered all explanatory variables at their means (29).

Let the P value of testing be P_INT. The presence of G×S will be declared if P_INT < (0.05/[3 × 5] = 0.0033, because three diabetes-related traits and five smoking measurements were investigated. This significance level for GRS analyses was determined a priori and was used throughout this study.

One of the mechanisms behind G×S is epigenetics (30). Cigarette consumption has been found to be linked to differential DNA methylation in some diabetes susceptibility genes, such as KCNQ1 (potassium voltage-gated channel subfamily Q member 1) (31).

When one is smoking, carbon monoxide displaces oxygen from the person’s hemoglobin. This will decrease the amount of oxygen available for delivery to many tissues. Smoking can therefore affect one’s entire body including cardiovascular and circulatory systems (32). TWB also released blood DNA methylation data of 2,091 TWB subjects. These individuals were randomly selected from TWB1 (833 subjects) and TWB2 (1,258 subjects). After GRS-smoking interaction analysis, we also explored the associations between smoking and DNA methylation of diabetes susceptibility genes.

Individual-level TWB data are available upon application to TWB (https://www.twbiobank.org.tw/new_web/). Our application for the TWB data was approved in February 2020 with the accession number TWBR10810-07.

Our discovery cohort contained 25,460 TWB1 subjects, where 3,005 (11.8%) were (active) smokers and 22,455 were nonsmokers. The replication cohort included 58,774 TWB2 subjects, 5,173 (8.8%) were (active) smokers and 53,601 were nonsmokers. Table 1 presents basic characteristics of these two cohorts. The average FG and HbA_1c levels are higher in smokers than in nonsmokers. Model 1 was fitted to assess whether smoking was significantly associated with FG, HbA_1c, and diabetes, while considering the covariates. As shown by Table 2, male, aging, and a larger BMI, were associated with increased FG and HbA_1c and risk of diabetes.

Because the FG and HbA_1c term was natural log transformed, is shown to represent the percent change in FG and HbA_1c that was associated with active smoking status, while adjusting for sex, age, BMI, drinking status, regular exercise, educational attainment, and the first 10 PCs. Both our discovery cohort (TWB1) and replication cohort (TWB2) showed that compared with nonsmokers, active smokers have an odds ratio (OR) of 1.36 for diabetes (95% CI for TWB1 1.19–1.57; 95% CI for TWB2 1.23–1.51) (Table 2).

On the basis of model 1, we further analyzed the four continuous smoking measurements, respectively (Table 3). For example, TWB1 (TWB2) showed that, subjects consuming one more pack of cigarettes per day have an OR of 1.51 (1.41) for diabetes (95% CI for TWB1 1.32–1.73; 95% CI for TWB2 1.27–1.56). The detailed results of model 1 for the four continuous smoking measurements can be found in Supplementary Tables 1–4.

Analyzing TWB1 according to model 2, a total of 16, 12, and 6 SNPs were identified to have in FG, HbA_1c, and diabetes analyses, respectively. The information of these SNPs was listed in Supplementary Tables 5–7. A total of 5 out of the 16 FG-associated SNPs, 4 out of the 12 HbA_1c-associated SNPs, and 2 out of the 6 diabetes-associated SNPs were also identified by another trait. The GRS of each TWB2 individual was then calculated based on these 16, 12, and 6 SNPs, with weights obtained from the TWB1 analysis (i.e., β_SNP from model 2).

Table 4 presents the results of model 4 when the smoking measurement is active smoking status. FG (or HbA_1c) was natural log transformed, and therefore, represents that each 1 SD increase in GRS was associated with a 1.05% higher FG (or 0.51% higher HbA_1c) in active smokers than in nonsmokers (⁠⁠; ⁠). Each 1 SD increase in GRS was associated with a 1.16 times OR (95% CI 1.05–1.28) for diabetes in active smokers than in nonsmokers ⁠. After analyzing the four continuous smoking measurements sequentially according to model 4, we found that all smoking measurements, except hours as a passive smoker per week, were associated with the exacerbation of the genetic risk of diabetes (Table 5).

For example, each 1 SD increase in GRS was associated with a 1.68% higher FG (or 0.69% higher HbA_1c) in subjects with one more pack of cigarettes per day (⁠⁠; ⁠). The detailed results of model 4 for the four continuous smoking measurements can be found in Supplementary Tables 8–11.

Figure 1 shows the average of FG and HbA_1c and the prevalence of diabetes stratified by smoking status and the quintiles of the FG, HbA_1c, and diabetes GRS. The GRS effects on FG, HbA_1c, and diabetes were larger in smokers than in nonsmokers. Supplementary Figures 1–4 present similar illustrations for the other four smoking measurements. Except for passive smoking (Supplementary Fig. 4), all smoking measurements exhibited interactions with GRS on FG, HbA_1c, and diabetes. This is in line with the result shown in Table 5.

Among the five smoking measurements, only active smoking status and years as a smoker presented interactions with GRS on diabetes under (Table 5). The power to identify diabetes-associated SNPs was low because only 2,207 (8.7%) among the 25,460 TWB1 subjects had diabetes. As a result, we merely detected 6 diabetes-associated SNPs at the significance level of (Supplementary Table 7). The GRS constructed by these six SNPs may not be able to well represent the genetic susceptibility to diabetes. Moreover, only 5,308 (9.0%) among the 58,774 TWB2 subjects had diabetes, the power to detect GRS-smoking interactions was limited. Therefore, none of the five smoking measurements were found to be associated with the exacerbation of the genetic risk of diabetes under (the last column of Table 5).

Although the response variable in model 4 is the natural log transformed term for FG and HbA_1c, these significant findings in GRS-smoking interactions are not scale dependent. Supplementary Table 12 shows that these significant results are still replicable when the response variable is FG and HbA_1c without natural log transformation.

Multicollinearity between variables in model 4 has been evaluated via the variance inflation factor (VIF), computed by the car package (https://cran.r-project.org/web/packages/car/). The VIFs under all models were acceptable (smaller than 10), except the model to assess the interaction between GRS and hours as a passive smoker per week. To reduce the multicollinearity between variables in model 4, we further assessed GRS-smoking interactions without controlling for the other interaction terms (i.e., no ⁠). The VIFs under these simpler models were all smaller than 5. Moreover, Supplementary Table 13 shows that the results from simpler models were similar to Table 5.

We then assessed GRS-smoking interactions by sex. Supplementary Table 14 shows that the significant GRS-smoking interactions were mostly driven by the male group. This is because 88.1% and 77.4% smokers were males in TWB1 and TWB2, respectively. The detection of GRS-smoking interactions in females could be hampered by the small sample size of female smokers.

Four SNPs (rs2399794, rs7896600, rs1174605899, and rs11257655) in or near the CDC123 (cell division cycle 123) gene were identified to be associated with FG, HbA_1c, and diabetes. They presented interactions with most smoking measurements except passive smoking (most ⁠) (Supplementary Tables 5–7). The CDC123 gene has been found to be associated with the dysfunction of pancreatic β-cells (33). The inability of pancreatic β-cells to secrete adequate levels of insulin is a major cause of diabetes (34).

Two SNPs (rs2233580 and rs61342118) in or near the PAX4 (paired box 4) gene were identified to be associated with FG, HbA_1c, and diabetes. They presented interactions with some smoking measurements except passive smoking (some ⁠) (Supplementary Tables 5–7). The PAX4 gene is necessary for the survival and proliferation of pancreatic β-cells (35). This gene has been found to be hypermethylated in patients with diabetes (36).

Four SNPs (rs163177, rs60808706, rs11024175, and rs163184) in KCNQ1 were identified to be associated with FG, HbA_1c, and diabetes. SNP rs11024175 was found to interact with all the five smoking measurements on HbA_1c  (Supplementary Table 6). KCNQ1 mRNAs were highly expressed in adrenal tissues (37). Disorders of the adrenal cortex can result in glucose intolerance and diabetes (38). Consistent with our finding, KCNQ1-by-smoking interaction has been reported by a study in Han Chinese subjects (39).

The 2,091 individuals with methylation measures were randomly selected from TWB1 (833 subjects) and TWB2 (1,258 subjects). Their basic characteristics (Supplementary Table 15) are similar to those of the whole TWB (Table 1). We annotated CpG sites available on the Illumina Infinium MethylationEPIC BeadChip to the abovementioned three genes. A total of 72, 84, and 629 CpG sites are within or near CDC123, PAX4, and KCNQ1, respectively (“near” indicates 50 kb in the 3′ and 5′ regions outside the gene boundary). The methylation percentage of a CpG site was reported as a β-value ranging from 0 (no methylation) to 1 (full methylation). The β-value of each CpG site was regressed on the number of packs smoked per day, while adjusting for age, sex, and BMI. These three covariates were also adjusted in a related study (31). Although packs smoked per day served as the smoking factor, active smoking status provided similar results.

Packs smoked per day was associated with differential DNA methylation of CDC123 (cg06335123, P = 0.00025 < 0.05/72) and KCNQ1 (cg26963277, ⁠; cg01744331, ⁠; cg16556677, < 0.05/629). More packs smoked per day, aging, and a larger BMI are associated with decreased levels of DNA methylation at these four sites (Supplementary Table 16). The three CpG sites in KCNQ1 (cg26963277, cg01744331, and cg16556677) have been reported to be associated with smoking in the Rotterdam study (31). Our finding from Han Chinese is in line with that study (31).

Through its link with differential DNA methylation, smoking can modulate gene expressions of diabetes susceptibility genes and lead to more damage to people who are more genetically predisposed to diabetes. However, we have not observed the association between smoking and DNA methylation of PAX4. The mechanism behind PAX4-by-smoking interaction needs further investigation.

A limitation is that our DNA methylation measures were made in peripheral blood, and the more relevant tissue is probably pancreatic β-cell. Although we here found that active smoking is associated with methylation at CpG sites in some diabetes susceptibility genes, additional work is necessary to determine if this mediates the smoking-genotype interaction.

Smoking is associated with insulin resistance, inflammation, and dyslipidemia (40). Consistent with our finding, the KCNQ1-by-smoking interaction has been reported by a study in Han Chinese patients (39). Moreover, Wu et al. (15) recently identified interactions between smoking status and five SNPs at or near four genes (TCF7L2, CUBN, C2orf63, and FBN1) on the risk of diabetes, in subjects of European and African ancestry. Two out of the five SNP-smoking interactions can be replicated in our TWB2 HbA_1c and diabetes analysis, at the nominal significance level of 0.05 (Supplementary Table 17). Both of the variants locate at the TCF7L2 (transcription factor 7 like 2) gene. TCF7L2 is a diabetes-associated gene identified in subjects of European ancestry (41), but the variants in this gene are not associated with FG, HbA_1c, and diabetes in TWB1. Therefore, they were not selected to form our GRS values. Variants in TCF7L2 are associated with pancreatic β-cell function (42).

For diabetes, Langenberg et al. (1) showed no interactions between genetic risk and physical activity or dietary habits. The recent study by Said et al. (14), which was based on the UK Biobank, further found no significant interactions between lifestyle and GRS on diabetes. However, the lifestyle variable used by Said et al. (14) is a composite measure of active smoking status, BMI, and physical activity. It remains unclear whether smoking alone may modify the genetic predisposition to diabetes. Moreover, the GRS used by Said et al. (14) for diabetes was calculated by 38 diabetes-associated SNPs with effect sizes retrieved from previously published GWASs (24–27). Most of these 38 SNPs were identified at the commonly used genome-wide significance level ⁠.

Individuals of those GWASs (24–27) were overwhelmingly of European descent. A SNP in the abovementioned TCF7L2 gene, rs7903146, is also among the list of the 38 SNPs (Supplementary Tables 18–20). One of the 38 SNPs, rs13266634 (in the SLC30A8 gene), is also among our six diabetes-associated SNPs (Supplementary Table 7). Moreover, the 38 SNPs included variants in or near CDC123 and KCNQ1, which were also identified to be associated with FG, HbA_1c, and diabetes according to our discovery cohort (TWB1), as listed in Supplementary Tables 5–7.

We then found three smoking factors (active smoking status, the number of pack-years, and years as a smoker) were associated with an exacerbation of EuGRS on both FG and HbA_1c (Supplementary Tables 21 and 22), at the nominal significance of 0.05 ⁠. The EuGRS × packs smoked per day interaction presented a borderline significance on FG ⁠. Similar with the results based on TWB1-GRS, EuGRS × hours as a passive smoker per week interaction was not significant on FG or HbA_1c (Supplementary Table 22).

We also analyzed diabetes as a dichotomous trait. Through logistic regression models, we did not detect any significant EuGRS-smoking interactions. In addition to low statistical power (<10% subjects with diabetes), this negative finding may also result from the lack of transferability of a European-derived GRS to TWB. Comparing Table 4 with Supplementary Table 21, we can see that TWB1-GRSs are more significant (smaller P values) and more predictive (larger R²) than EuGRS for all the three traits, despite fewer SNPs used for TWB1-GRSs. The detection of EuGRS-smoking interactions could be hampered by the inferior transferability of EuGRS to TWB.

Supplementary Table 23 shows that the results of EuGRS-smoking interactions are still replicable when the response variable is FG and HbA_1c without natural log transformation. When assessing EuGRS-smoking interactions without controlling for the other interaction terms (i.e., no EuGRS × Covariate_v or Smoking × Covariate_v, v = 1, ..., 16), Supplementary Table 24 shows that the results from simpler models were similar to Supplementary Table 22. Moreover, in line with the results from TWB1-GRS, Supplementary Table 25 also shows that the significant EuGRS-smoking interactions were mostly driven by the male group.

Diabetes is a growing health crisis around the world. According to the Centers for Disease Control and Prevention and many previous studies, smoking increases inflammation in the body (43) and also causes oxidative stress (44). Both inflammation and oxidative stress have been shown to be linked to an increased risk of diabetes (45). Oxidative stress is an imbalance between oxidants and antioxidants in favor of the oxidants (46), potentially modulating gene expressions (47) and leading to more damage to people who are more genetically predisposed to diabetes. Indeed, we here found smoking is associated with DNA methylation at KCNQ1 and CDC123, implying that smoking can play a role in regulating gene expressions of these diabetes susceptibility genes.

According to previous studies (48), the amount of nicotine (the main chemical in cigarettes) absorbed by a passive smoker was between one-tenth and one-third of the amount in a cigarette. Therefore, the concentration of nicotine in passive smokers is smaller than that in active smokers (48). This may explain why passive smoking does not significantly modulate the genetic predisposition to diabetes.

The harm of smoking is more impactful in subjects who are more genetically predisposed to diabetes. This study shows that active cigarette smoking is associated with an exacerbation of genetic risk of diabetes. Through constructing a GRS based on diabetes associated SNPs, it is worthwhile to investigate whether smoking may exacerbate the genetic susceptibility to diabetes, even for populations where smoking has not been found to be associated with diabetes.

Acknowledgments. The authors thank the editors and three anonymous reviewers for their insightful and constructive comments, Tsung-Hao Lee and Yu-Sheng Lee (Institute of Epidemiology and Preventive Medicine, National Taiwan University) for imputation of the TWB data, and Ya-Chin Lee (Institute of Epidemiology and Preventive Medicine, National Taiwan University) for assisting with the acquisition of the TWB data.

Funding. This study was supported by the Ministry of Science and Technology, Taiwan (grant number MOST 107-2314-B-002-195-MY3 to W.-Y.L.), and the acquisition of TWB data were supported by two MOST grants (MOST 107-2314-B-002-195-MY3 to W.-Y.L., MOST 102-2314-B-002-117-MY3 to P.-H.K.) and a collaboration grant (National Taiwan University Hospital grant number UN106-050 to P.-H.K.).

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. W.-Y.L. conceived the study design, developed the analysis tool, analyzed the TWB data, and wrote the manuscript. Y.-L.L., A.C.Y., S.-J.T., and P.-H.K. contributed to the writing of the manuscript. All authors provided the TWB data and reviewed the manuscript. W.-Y.L. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.