A genome-wide search for gene-by-obesity interaction loci of dyslipidemia in Koreans shows diverse genetic risk alleles

Dyslipidemia is a well-established risk factor for cardiovascular disease. Studies suggest that similar fat accumulation in a given population might result in different levels of dyslipidemia risk among individuals; for example, despite similar or leaner body composition compared with Caucasians, Asians of Korean descent experience a higher prevalence of dyslipidemia. These variations imply a possible role of gene-obesity interactions on lipid profiles. Genome-wide association studies have identified more than 500 loci regulating plasma lipids, but the interaction structure between genes and obesity traits remains unclear. We hypothesized that some loci modify the effects of obesity on dyslipidemia risk and analyzed extensive gene-environment interactions (GxEs) at genome-wide levels to search for replicated gene-obesity interactive single-nucleotide polymorphisms (SNPs). In four Korean cohorts (n=18,025), we identified and replicated 20 gene-obesity interactions, including novel variants ( SCN1A and SLC12A8 ) and known lipid-associated variants ( APOA5 , BUD13 , ZNF259 , and HMGCR ). When we estimated the additional heritability of dyslipidemia by considering GxEs, the gain was substantial for triglycerides (TGs) but mild for low-density lipoprotein cholesterol (LDL-C) and total cholesterol (Total-C); the interaction explained up to 18.7% of TG, 2.4% of LDL-C, and 1.9% of Total-C heritability associated with waist-hip ratio. Our findings suggest that some individuals are prone to develop abnormal lipid profiles, particularly with regard to TGs, even with slight increases in obesity indices; ethnic diversities in the risk alleles might partly explain the differential dyslipidemia risk between populations. Research about these interacting variables may facilitate knowledge-based approaches to personalize health guidelines according to individual genetic profiles.

10 2p(1-p)(log(OR)) 2 , where p is the MAF of a variant and OR is the estimated odds ratio from a logistic regression model for marginal associations (47). The contribution of each GxE marker was estimated using 2p(1-p)(log(ORG)) 2 /VP+2ep((1-p)+2p(e-1) 2 )(log(ORGxE)) 2 /VP. In this equation, e is the prevalence of an environmental factor, VP is the phenotypic variance, and ORG and ORGxE are estimated additive and geneobesity interactive ORs from a logistic regression model for GxEs. We used GenABEL, the R package for genome-wide association analyses (48), to estimate the total heritability of dyslipidemia from the Healthy Twin Study, a family-based cohort study in Korea (Supplemental Table S2.e) (49, 50). GCTA, an analysis tool for genome-wide complex traits (51), was also used to estimate the SNP-based heritability attributable to all GWAS variants genotyped on a microarray. To transform the estimate of variance explained on the observed scale to that on the underlying scale, we assessed the prevalence of dyslipidemia using data from the Korean National Health and Nutrition Examination Survey (KNHANES) (52). Table 1 shows the baseline characteristics of the participants in each Korean genome cohort. We observed the age, sex, obesity-related traits, and age-and sex-standardized plasma lipid levels of the cohorts; all features were stratified by obesity status into subgroups based on BMI, WC, and WHR (Supplemental Table   S1 and S2). We focused on the adjusted lipid concentrations to assess the trends of lipids in each obesity and abdominal obesity subgroup. As expected, age-and sex-adjusted lipid levels significantly worsened as the degree of obesity status increased in the combined Korean cohort (Supplemental Figure S1).

Results
We identified 55 SNPs showing genome-wide significant GxE effects on the risk of abnormal lipid profiles with at least one of the six obesity traits (Supplemental Table S3). By conducting LD clumping based on the genetic contribution to the risk of dyslipidemia, we detected 20 gene-obesity interactions due to novel SNPs near SCN1A and SLC12A8 and to lipid-associated SNPs near APOA5, BUD13, ZNF259, and HMGCR that were reported in previous GWASs. Table 2 shows the marginal and gene-obesity interactive effects of the newly identified variants on the risk of dyslipidemia; we summarized the novel GxEs according to the discriminators of obesity traits such as BMI, WC, and WHR. Figure 1 (Supplemental Table S4) describes the risk of abnormal lipid profiles for each genetic and environmental factor; we estimated the OR as the ratio of the probability of dyslipidemia occurring in each exposed group (G≠0 or E≠0) to the probability in a non-exposed group (G=0 and E=0).
We identified three novel SNPs interacting with obesity traits to modify the risk of abnormal elevation of Total-C: rs2878417, rs7702895, and rs7733436. In particular, COL4A3BP exhibited synergistic effects with BMI and WC on the risk of abnormalities in Total-C. For the interplay between HMGCR and WHR, the marginal odds ratio (ORD) was 0.81 (95% CI, 0.78-0.84); ORG and ORGxE were 0.72 (95% CI, 0.68-0.77) and 1.22 (95% CI, 1.13-1.30), respectively. As shown in Figure 1.a (Supplemental Table S4.a), the multiplicative effect of abdominal obesity was 1.12 (95% CI, 0.96-1.31) for individuals with two wild-type alleles at rs7702895. The magnitude of the effect of abdominal obesity, however, increased with the number of minor alleles, with values of 1.46 (95% CI, 1.28-1.67) for heterozygous and 1.57 (95% CI, 1.26-1.95) for homozygous minor alleles.   Table S4.e) describes the gene-obesity interactive effect on the risk of abnormal HDL-C reduction; the multiplicative effect for common homozygous or heterozygous and rare homozygous genotypes was 1.42 (95% CI, 1.22-1.65), 1.99 (95% CI, 1.57-2.52), and 6.24 (95% CI, 4.03-9.64), respectively.  Table   S4.b), the multiplicative effect of overweight class 1 was 1.00 (95% CI, 0.54-1.82) for rare homozygous genotypes. For common homozygous or heterozygous genotypes, on the other hand, obesity acted as a risk factor for abnormalities in LDL-C; the multiplicative effect was 1.82 (95% CI, 1.61-2.06) and 1.34 (95% CI, 1.12-1.61), respectively. with our previous findings. On the other hand, we could not detect any interactions of GxE markers with BMI; only the variants, previously detected with multiple analytical methods for testing GxEs (Table 2), were consistently identified. Similarly, we identified only one GxE SNP of dyslipidemia, located on APOA5, by using the alternative definition of obesity, the highest quintile of BMI or WC or WHR (Supplemental Table S8); we could not find any loci interacting with BMI or WC. Table 3 shows the contributions of marginal associations and gene-obesity interactions to abnormal lipid profiles; we present the proportion of total heritability for each lipid explained by GWAS-identified SNPs, novel GxE loci, and the combined set of both lipid-associated and gene-obesity interactive variants (total genetic impact). The total and SNP-based heritabilities of the risk of abnormalities in Total-C were approximately 35.5% and 17.7-24.9%, respectively, after adjusting the risk of dyslipidemia by age, age 2 , and sex. The genetic contributions increased when we considered both marginal associations and geneobesity interactions, with differences between the GWAS-identified and total genetic impact of 1.1-1.9%.
The total and SNP-based heritabilities of the risk of LDL-C abnormalities, on the other hand, were approximately 31.7% and 17.2-25.6%, respectively. For each obesity trait, the total genetic contributions including gene-obesity interactions to the risk of dyslipidemia were 0.9-2.4% higher than the marginal impact due only to direct associations.
The contributions of the combined set of both GWAS-identified and GxE variants were markedly higher when several independent gene-obesity interactive loci were present for each pair of lipid traits and environmental factors. Figure 3.a, 3.b, and 3.c (Table 3) present the risk of abnormal elevation of TG.
Genetic factors accounted for approximately 38.3% of the total variance of the risk of abnormalities in TG after adjusting the risk by age, age 2 , and sex. Genetic markers located on the genome-wide dense SNP microarray accounted for 18.4-26.4% of the overall variance for the risk of hypertriglyceridemia.
Approximately 36.6% of the total heritability was due to 40 independent GWAS-identified SNPs only; the genetic contribution increased to 47.1% when we considered the interactions of APOA5 or BUD13 with WC. Similarly, the total genetic impact increased from 39.3% to 58.0% when we considered both marginal associations and newly found genetic interactions attributable to WHR. For Caucasians, the additional TG heritability due to the interactions of GxE variants with WC or WHR was 5.8% and 9.1%; the gain was 10.6% and 18.7% for Koreans, respectively.
The genetic contributions to the risk of abnormal elevation of Remnant-C are described in Figure 3.d (Table   3). Genetic factors explained approximately 48.6% of the total variance after adjusting for age, age 2 , and sex. Genotyped loci on the SNP microarray accounted for 11.3-14.2% of the overall variance for Remnant-C. Approximately 38.5% of the total heritability was explained by 59 independent GWAS-identified SNPs only. When both marginal associations and interactions of APOA5 or BUD13 with WHR that modify the risk of abnormal Remnant-C were considered, the genetic contribution increased to 47.8%; the difference between marginal and total genetic impact was 9.3%. For Caucasians, on the other hand, the additional heritability for Remnant-C was just 5.1%.

Discussion
One of the main purposes of human genome studies is to personalize treatment and health guidelines according to an individual's genetic constitution. GWISs are approaches intended for achieving this end, particularly when genetic loci interacting with modifiable risk factors are examined at a genome-wide level.
Such studies permit the identification of higher-or lower-risk individuals depending on changes in known risk factors. In this study, we identified novel and known genes interacting with obesity indices to modify the risk of dyslipidemia. We also replicated our findings using independent genome cohorts and assessed how much phenotypic variance or heritability was additionally explained by considering the gene-obesity interactions.
Our study focused on increasing power to detect gene-obesity interactions by applying a variety of strategies for testing GxEs. We carried out emerging exhaustive scans and two-step methods in parallel because each analytical model provided differential power to detect GxEs, mainly according to marginal genetic and GxE effects. We tested interactions of SNPs at a genome-wide scale with several obesity traits, including Koreanspecific parameters defined by additional ranges of BMI and WC. Besides, we adopted liberal cut-offs and stepwise penalties due to marginal p-values as well as the standard genome-wide significance level to find gene-obesity interactive loci influencing the risk of dyslipidemia. Type 1 errors are generally considered to be less problematic than possible underpowered findings (27); it is recommended to use multiple models for GxEs. Further verification can be done by replication and stratified analyses for candidate GxE regions.
Our findings reveal a genome-wide set of variants with a wide range of marginal effects on the risk of dyslipidemia. We identified novel GxE markers near SCN1A and SLC12A8 with little or no direct association with lipid parameters as well as gene-obesity interactions related to lipid-associated loci reported in previous GWASs on lipids: APOA5, BUD13, ZNF259, and HMGCR. We identified SCN1A and SLC12A8 through exhaustive CO analyses, while all other GxEs due to lipid-associated loci were detected using two-step methods due to the marginal effects of each locus in the first step. These trends are consistent with the results of an earlier simulation study of statistical power for GxE detection, which showed that exhaustive CO analysis is more powerful than other two-step methods when the marginal effects of genetic variants are small (43). To our knowledge, SCN1A and SLC12A8 have not been previously associated with any lipid parameters.
We replicated the novel findings in four Korean genome cohorts; one strength of using cohorts formulated on identical protocols is the ability to examine gene-obesity interactions with high-quality health outcomes, genetic and environmental factors. In addition, conducting meta-analyses with the independent Korean cohorts permitted the estimation of more precise effects of susceptibility loci interacting with obesity traits.
We also classified individuals in this study into three groups according to the number of risk alleles at GxE loci and compared the changes in lipid levels when BMI, WC, and WHR increased by one unit between the three groups. This comparison reconfirmed the identified gene-obesity interactions from different points of view, as the changes in lipids due to the elevation of obesity indices worsened as the number of risk alleles increased.
Although generating interesting findings, our approaches for testing gene-obesity interactive effects on lipid profiles are not free of limitations. Our study did not include GxEs due to loci marked by rare variants (MAF<0.01) and other essential obesity indices, such as body fat percentage and visceral fat level. We focused only on the GxE effects due to a set of common variants since our study populations did not include enough information for rare genetic variants. In addition, current analytical methods do not provide the GxE analyses using TG levels quantitatively (Supplemental Table S6). For GWISs, it is more important to use multiple analytical models as possible to generate a consistent and powerful result.
For categorical analyses, the cut-offs were decided by clinical guidelines for managing hyperlipidemia and obesity. For HDL-C, on the other hand, we used quantiles because the clinical cut-off points (HDL-C<1.03 mmol/L for males, 1.29 mmol/L for females) resulted in too many dyslipidemia cases (46.8%) in our study population; the arbitrary cut-offs could affect the results for interactions. To clarify the issue, we conducted GWISs using two different methods of categorization: by clinical guidelines and by the quantile distribution in our datasets. Compared with using quantiles to determine dyslipidemia and obesity traits, the categorical GxE analyses using clinical thresholds could find out a more extensive range of gene-obesity interactions (Supplemental Table S7 and S8). The clinical cut-off points, commonly accepted and ascertained by several epidemiological studies, were more appropriate to detect interactions at a genome-wide scale than quantiles dividing the study population into equal-sized bins for phenotypes.
Our ability to extend these novel findings from Korean populations to other ethnic groups is limited by differences in the MAFs of genetic markers, distributions of obesity traits, and prevalences of each lipid abnormality. Our results were estimated and reconfirmed in four independent Korean cohorts sharing phenotyping and genotyping protocols, and the identified gene-obesity interactions in the risk of dyslipidemia might not be supported when racial differences in lipid profiles and the distribution of genetic and environmental factors are considered. ZNF259 marked by rs2075291, for example, could be a useful therapeutic target for managing TG in Korean population; the ZNF259-WHR interactive impact on the risk of hypertriglyceridemia was 3.4%. This genetic variant, however, is not a suitable target for the other ethnic groups; the minor allele of rs2075291 is infrequent for South Asians, and too rare for Europeans, Americans, and Africans (Table 3) Table S4.