Adiposity significantly modifies genetic risk for dyslipidemia[S]

Recent genome-wide association studies have identified multiple loci robustly associated with plasma lipids, which also contribute to extreme lipid phenotypes. However, these common genetic variants explain <12% of variation in lipid traits. Adiposity is also an important determinant of plasma lipoproteins, particularly plasma TGs and HDL cholesterol (HDLc) concentrations. Thus, interactions between genes and clinical phenotypes may contribute to this unexplained heritability. We have applied a weighted genetic risk score (GRS) for both plasma TGs and HDLc in two large cohorts at the extremes of BMI. Both BMI and GRS were strongly associated with these lipid traits. A significant interaction between obese/lean status and GRS was noted for each of TG (PInteraction = 2.87 × 10−4) and HDLc (PInteraction = 1.05 × 10−3). These interactions were largely driven by SNPs tagging APOA5, glucokinase receptor (GCKR), and LPL for TG, and cholesteryl ester transfer protein (CETP), GalNAc-transferase (GALNT2), endothelial lipase (LIPG), and phospholipid transfer protein (PLTP) for HDLc. In contrast, the GRSLDL cholesterol × adiposity interaction was not significant. Sexual dimorphism was evident for the GRSHDL on HDLc in obese (PInteraction = 0.016) but not lean subjects. SNP by BMI interactions may provide biological insight into specific genetic associations and missing heritability.

novel, and several encompass genes not previously implicated in plasma lipid metabolism. Furthermore, these loci were shown to contribute not only to general variation in plasma lipids, but also to extreme lipid phenotypes ( 3 ). Notably, for TGs, individuals in the top quartile of the TG risk score were 44 times more likely to have hypertriglyceridemia as compared with individuals in the bottom quartile ( P = 4 × 10 Ϫ 28 ). For HDL cholesterol (HDLc), individuals in the top quartile of the risk score were four times more likely to have high HDLc as compared with those in the bottom quartile ( 1 ).
Although family-based association studies indicate that 40% to 60% of variation in plasma TG and HDLc is genetically based ( 4,5 ), the identifi ed loci explain <12% of variation in each of these lipid traits ( 1 ). Environmental and clinical factors including BMI, physical activity, and alcohol intake are also important determinants of plasma TG and HDLc ( 6 ).
Thus, interactions between genetic risk factors and clinical phenotypes may account for some of the unexplained heritability of plasma lipid traits. Here we have examined whether the effect of a weighted genetic risk score (GRS) on each of TG and HDLc is modifi ed by adiposity, as assessed by BMI. This study provides biological insight into specifi c genetic associations and may aid in the identifi cation of dyslipidemic subjects for whom weight loss is likely to be an important intervention.
Abstract Recent genome-wide association studies have identifi ed multiple loci robustly associated with plasma lipids, which also contribute to extreme lipid phenotypes. However, these common genetic variants explain <12% of variation in lipid traits. Adiposity is also an important determinant of plasma lipoproteins, particularly plasma TGs and HDL cholesterol (HDLc) concentrations. Thus, interactions between genes and clinical phenotypes may contribute to this unexplained heritability. We have applied a weighted genetic risk score (GRS) for both plasma TGs and HDLc in two large cohorts at the extremes of BMI. Both BMI and GRS were strongly associated with these lipid traits. A signifi cant interaction between obese/lean status and GRS was noted for each of TG ( P Interaction = 2.87 × 10 GRS SNPs were individually coded as 0, 1, or 2, according to the number of trait-increasing alleles at that particular SNP. To generate the GRS TG , 20 SNPs were analyzed in the population; to generate the GRS HDL , 34 SNPs were analyzed. To generate GRS for LDL cholesterol (GRS LDLc ), 11 SNPs were analyzed. Several SNPs for each trait failed to pass QC in our populations and were thus excluded from analysis. If a particular SNP failed QC in a particular subgroup, it was coded as missing in the total population. A weighted GRS ( Ŝ ) was constructed for each individual by taking a sum across SNPs of the number of reference alleles (0, 1, or 2) at that SNP and multiplying by the ␤ effect score of that allele. Thus, we defi ne G as an m-vector of coded markers (0, 1, or 2) and ␤ as the effect size at that allele defi ned by the Global Lipids Consortium ( 1,13,14 ).
After experimentation with various methodologies, we concluded that a weighted GRS outperforms allele counting or a merely additive model ( 9,13,15 ). GRSs were constructed in PLINK: whole-genome association analysis toolset ( 14 ). SNPs and corresponding effect sizes for each of TG and HDLc are provided in supplementary Table I. Effect sizes provided are for the primary trait only.

Statistical analysis
Individual post-QC genotyped SNPs were coded as 0, 1, or 2 according to the number of effect alleles present, and a weighted GRS was constructed for each individual according to the previously described procedure for each of TG and HDLc. Multiple general linear regression models (GLMs) were used to test for the association between genotypes and HDLc and TGs. Data were adjusted for age, sex, and age 2 . Response data were broken down into lean, obese and normal range categories in order to investigate the effect of genetic risk across the BMI spectrum. Each SNP was tested for associations to phenotype separately from the GRS using GLMs, and interaction scores were constructed for SNP × obese/ lean status and SNP × sex by including an interaction term in the respective models. The same covariates, which were used to analyze the data, were also controlled for when determining SNP × obese/lean status and SNP × sex interaction terms. Data were further stratifi ed by gender. All analyses were conducted in PLINK ( 14 ) and R version 3.0.0 (http://www.r-project.org/).

RESULTS
The general characteristics of obese and lean subjects in each of the two main cohorts are shown in Table 1 . Within the OBLE and CAD-C cohorts, subjects were well matched for age and sex. The OBLE cohort was younger and exhibited greater extremes of BMI [mean 43.1 ± 0.3 (obese); 20.3 ± 0.1 kg/m 2 (lean)] as compared with the CAD-C group [mean BMI 34.6 ± 0.2 (obese); 21.3 ± 0.1 kg/m 2 (lean)].

Study subjects
Subjects with a BMI у 30 kg/m 2 were defi ned as obese, those with a BMI р 23 kg/m 2 as lean, and intermediate subjects (30 kg/ m 2 у BMI у 23 kg/m 2 ) as normal range. The BMI cutoff of р 23 for the lean subgroup is below the 25th percentile for the majority of individuals studied. Two cohorts were studied.
Obese versus lean. Obese, unrelated subjects of strictly European ancestry were recruited from the University of Ottawa Weight Management Clinic. Obese individuals displayed a BMI of >35 kg/m 2 and a history of at least 10 years of adult obesity with no medical or psychiatric predisposing factors. Unrelated lean subjects were recruited from the Ottawa community. These healthy individuals had a lifelong BMI of less than the 25th percentile for sex and age, and no medical or psychiatric conditions affecting body weight ( 7,8 ). Body weight was measured using a Tanita electronic scale to the nearest 0.3 kg. BMI was defi ned as weight in kilograms divided by height in meters squared (kg/m 2 ). Height was measured to the nearest 0.5 cm. Plasma lipid fractions were measured using standard procedures. For coronary artery disease controls (CAD-C) subjects on lipid modifying medication, written documentation of pretreatment plasma lipids was obtained from the primary care physician and used for these analyses. These data were not available for 6.4% of the CAD-C subjects, none of whom were treated with a fi brate or niacin. In the obese versus lean (OBLE) cohort, 2.6% of lean and 14.8% of obese subjects were on low-to moderate-dose statin therapy, not expected to have major effects on TG or HDLc. The study was approved by the Human Ethics Experimentation Committees of the University of Ottawa Heart Institute and the Ottawa Hospital and written informed consent was obtained from all subjects .

CAD-C.
Details of the CAD-C cohorts have been previously described ( 9 ). Briefl y, CAD-C included healthy controls recruited as part of the Ottawa Heart Genomics Study in collaboration with the Cleveland Clinic Gene Bank (OHGS_A and OHGS_CCGB_B). These subsets were combined together to form a single CAD-C sample. Subjects were collected under human research protocols approved by their respective committees.

Genotyping and imputation
SNP genotyping of the OBLE and CAD-C cohorts was performed on Affymetrix 6.0 or 500K Arrays at the University of Ottawa Heart Institute using the standard protocol recommended by the manufacturer and processed as described ( 10,11 ). Imputation was performed using IMPUTE2 and the August 2009 1000 Genomes European reference panel ( 12 ). After imputation, ‫ف‬ 5.5 M SNPs passed post-quality control (QC) measures (info >0.5, Hardy Weinberg Equilibrium >1e-6 , missing <10%).

Sex × lipid trait interactions
Next, we investigated whether GRS effects differed by sex. Of note, sex did not signifi cantly infl uence the effect of the GRS on any trait in the whole population (TG: P Interaction = 0.0925; HDLc: P Interaction = 0.0868; LDL: P Interaction = 0.189). However, for GRS HDLc on HDLc, there was a signifi cant interaction with sex in the obese ( P Interaction = 0.016) but not the lean ( P Interaction = 0.369) population. A sex dimorphic effect by obese/lean stratifi cation was not found for the other lipid traits. Further analysis of individual SNPs failed to identify signifi cant interaction terms in the whole Because obesity status signifi cantly infl uenced the clinical expression of these lipid trait loci, we determined the explained variance ( R 2 ) of the GRS TG and GRS HDLc in obese versus lean subjects. For GRS TG on TG, R 2 = 0.0614 for obese versus R 2 = 0.0345 for lean subjects, a 2-fold difference. An opposite trend was observed for GRS HDLc (based on HDLcraising alleles) on HDLc, R 2 = 0.0790 for lean versus R 2 = 0.0533 for obese. In contrast, for the GRS LDLc on LDLc, explained variance was only slightly higher in the obese ( R 2 = 0.0215) versus lean ( R 2 = 0.0172) populations. We next examined the individual SNPs included in the GRS TG and GRS HDLc . Three TG SNPs ( APOA5 , GCKR , and LPL ) and four HDLc SNPs ( CETP , LIPG , GALNT2 , and PLTP ) were found to have a signifi cant obese/lean status × SNP effect interaction term at a false discovery rate (FDR) of 20% ( Table 3 ). LPL and APOA5 achieved a 10% FDR for TG, and CETP , GALNT2 , and LIPG reached a 10% FDR for HDLc. However, at a 5% FDR, only LPL and CETP were signifi cant. Statistical correction for multiple testing was achieved by ordering each tested SNP from least to greatest P Interaction value. The largest interaction term that was less than the P FDR [i.e., the ratio of the position of the Fig. 1. TG residuals compared with GRS stratifi ed by lean versus obese status. Signifi cantly differently slope coeffi cients with 95% CIs are displayed, demonstrating a signifi cant interaction between obesity status and a GRS. The rate of increased TG residuals for an increased predisposition is displayed for obese (broken line) and lean (solid line) individuals. Increased risk in obese individuals corresponds to an increased expression of lipid levels above what would normally be expected. This dimorphic effect was dependent on three SNPs tagging APOA5 , glucokinase receptor ( GCKR ), and LPL , not before observed to have adipositydependent dimorphic effects.

DISCUSSION
Lifestyle and clinical factors may modify genetic risk. For example, the effect of a GRS on BMI was found to be signifi cantly attenuated in physically active versus sedentary individuals ( 17 ). To explore the effects of adiposity on genetic risk for dyslipidemia, we have utilized a GRS constructed from loci previously reported by the Global Lipids Consortium. We demonstrate that obesity status signifi -population for either TG or HDLc. However, one sexually dimorphic locus for HDLc was found in each of the lean (rs4846914 tagging GALNT2 ) and obese [rs605066 tagging Cbp/p300-interacting transactivator, with Glu/Asp-rich carboxy-terminal domain, 2 ( CITED2 )] populations. However, after correction for multiple testing, these loci were only nominally signifi cant (FDR = 15%) ( Table 4 ). More complete SNP × sex interaction data may be found in supplementary Table III.  Although we lack the statistical power necessary to detect the individual effects of all loci, we identifi ed seven novel loci not previously reported to have obesity-related dimorphic effects. SNPs tagging APOA5 ( P Interaction = 8.87 × 10 ) showed interaction with obese/lean status for TG. These encompass genes encoding proteins altering both hepatic TG synthesis and peripheral lipolysis. The GCKR gene product, the glucokinase regulatory protein, regulates glucokinase (GCK) activity competitively with respect to the substrate glucose, inhibiting GCK activity. Hepatic GCK activity enhances glycolytic fl ux, promoting hepatic glucose metabolism and increasing malonyl CoA availability, a major substrate for de novo hepatic lipogenesis ( 18 ). LPL and APOA5 encode major determinants of peripheral lipolysis of TG-rich lipoproteins, LPL and ApoA5, the latter a regulator of LPL activity ( 19 ). The effect sizes of the previously discussed TG loci were among the highest in this study ( APOA5 ␤ = 16.95, GCKR ␤ = 8.76, and LPL ␤ = -13.64) and not surprisingly were responsible for the signifi cant obese/lean status × GRS interaction. Consistently, in a Filipino population the APOA5 effect on plasma TG levels was found to be modifi ed by waist circumference ( 20 ), another measure of adiposity.
In a second stage, we performed a sex-stratifi ed analysis. The effect of neither weighted GRS TG nor GRS HDLc was found to be signifi cantly different for males versus females for the population as a whole. Importantly, sexual dimorphism for genetic effects on HDLc was entirely driven by the obese subjects ( P Interaction = 0.016) and was not evident in the lean ( P Interaction = 0.914) or all ( P Interaction = 0.0868) cantly alters the effect of genetic variants associated with increased TGs as well as those associated with higher levels of HDLc, but not LDLc.
For TG, the effect size ( ␤ ) of a weighted GRS TG in the obese population was nearly double that of the lean population ( ␤ = 0.480 vs. 0.261). As shown in Fig. 1 , for any GRS TG , plasma TG levels are greater for obese versus lean subjects. This is not surprising given known effects of substrate availability on hepatic TG synthesis; in obese individuals, the effect of nutrient excess outweighs the effect of known genetic variants at any GRS TG . In both obese and lean individuals, GRS TG associates with higher TGs, but the slope of the line for GRS TG versus TG differs for obese as compared with lean yielding a signifi cant interaction coeffi cient. Overall, the variance in plasma TG concentrations explained by the GRS TG was 6.14% for the obese subjects, nearly double that found for the lean population (3.45%).
In contrast to TGs, the effect of a GRS HDL , consisting of HDLc-raising alleles, on HDLc was greater for the lean ( ␤ = 2.347) than the obese ( ␤ = 1.466) population. For HDLc, it is important to note that we created a GRS HDL composed of HDLc-raising alleles (a genetic protective score for HDL). As shown, obese individuals have higher circulating levels of TG-rich lipoproteins, leading to TG enrichment of HDL and more rapid HDL clearance. Thus, as shown in Fig. 2 , it is likely that the metabolic effect of hypertriglyceridemia acts to attenuate the effect of HDL-raising alleles, for example near genes encoding CETP, LIPG, and PLTP . The GRS HDL for HDLc explained 7.89% of HDLc variation in the lean versus 5.33% in the obese subjects.
Thus, the genetic risk for hypertriglyceridemia is signifi cantly worsened by the obese state, whereas the beneficial effect of HDLc-raising genetic variants is attenuated. These data demonstrate that the gene × adiposity interaction contributes to part of the hitherto unexplained genetic variance in plasma lipids levels.
Here, we utilized an aggregated, weighted risk score rather than the more common allele counting method. In the past, allele counting, also known as an additive model, has been used due to a lack of well-established effect sizes ( 14 ). However, a weighted, aggregated risk score has been shown to improve power ( 7, 10 ). We did not perform receiver operating characteristic area-under-curve analysis because hypertriglyceridemia (high TG) and hypoalphalipoproteinemia (low HDL) are defi ned by age-and sex-dependent quantiles. groups. Obese men showed an attenuated increase in HDLc in response to GRS HDLc as compared with women ( Fig. 3 ). Loci in each subpopulation ( CITED2 for obese and GALNT2 for lean) were found to be dimorphic ( Table  4 ). However, after correction for multiple testing, these remained only nominally signifi cant (FDR <15%), thus requiring confi rmation in additional populations.
In summary, we have created weighted GRSs for each of TG and HDLc based on loci identifi ed by the Global Lipids Consortium and tested effects in separate large, welldefi ned obese and lean populations; thus, our results are without discovery bias. Neither GRS TG nor GRS HDLc showed an association with adiposity (BMI) per se. Here we demonstrate convincing gene-adiposity trait interactions. Notably, lean subjects have an ‫ف‬ 50% reduction in the genetic predisposition for increased TGs and an ‫ف‬ 35% greater response to HDLc-raising alleles, as compared with obese subjects. These effects are mainly driven by SNPs tagging APOA5 , GCKR , and LPL for TG, and CETP , LIPG , GALNT2 , and PLTP for HDLc. We also report sexual dimorphism for genetic effects on HDLc that is confi ned to the obese group of subjects. These fi ndings demonstrate that obese individuals are more susceptible to genetic risk for dyslipidemia. SNP by BMI interactions may provide biological insight into specifi c genetic associations and missing heritability. Fig. 3. Regression coeffi cients for HDLc (mM) for male versus female subjects stratifi ed by lean versus obese status. Data are shown for males and females stratifi ed by adiposity, and bars represent SE. In the lean population, women and men display a similar response to GRS HDLc . In contrast, obese men demonstrate an attenuated effect of GRS HDLc as compared with obese women. One locus was found to be exhibit sexually dimorphic effects in each of obese (rs4846914 tagging GALNT2 ) and lean (rs605066 tagging CITED2 ) populations as shown in Table 4 .