Introduction

Coronary artery disease (CAD) is a major cause of morbidity and mortality; much international effort has been expended to detect risk factors, both heritable and environmental.1 One of these, obesity, is associated with CAD, but it is not clear whether this association is directly causal, is due to confounding by conventional risk factors such as diabetes mellitus, or is due to reverse causation. Genome-wide association studies (GWAS) for body mass index (BMI) have identified multiple obesity risk alleles. Recent studies have demonstrated an association between carrier status for fat mass and obesity-related gene (FTO) risk alleles and cardiovascular disease, with un-replicated results implicating FTO plus a small number of other BMI risk SNPs as part of a genetic risk score (GRS).2, 3, 4, 5 Here, we have carried out a Mendelian randomization study to determine the effects of a GRSBMI, incorporating a much larger set of 35 BMI risk alleles, on CAD risk in subjects without diabetes mellitus, a major obesity-associated CAD risk factor, and with adjustment for traditional risk factors (TRFs). Our findings support the hypothesis that obesity is a causal risk factor for CAD and show that a GRSBMI improves the predictive value for CAD beyond that of a single measurement of BMI.

Methods

Study subjects

Details of the cohorts have been previously described.6 Briefly, the participants are part of four CAD case–control cohorts, recruited at the University of Ottawa Heart Institute (UOHI) Lipid Clinic, catheterization lab, or as part of the Cleveland Clinic Gene Bank study. Cases had a history of at least one of: myocardial infarction, coronary artery bypass grafting, percutaneous coronary intervention, or a coronary angiogram or computed tomography angiogram demonstrating a stenosis of at least 50% in at least one epicardial artery. Controls were either minimally burdened with disease (<30% stenosis in any major coronary artery) or asymptomatic for ischemic cardiovascular disease. Cases were ≤55 years of age for men and ≤65 years of age for women at onset, while controls were ≥65 years of age for men and ≥70 years of age for women. Subjects with diabetes mellitus were excluded. Subjects were collected under human research protocols approved by their respective committees.

Genotyping and imputation

SNP genotyping of the cohorts was performed on Affymetrix 6.0 Arrays and Affymetrix 500k Arrays (Santa Clara, CA, USA) at the University of Ottawa Heart Institute (UOHI) using the standard protocol recommended by the manufacturer and processed as described.6, 7 Imputation was performed using IMPUTE28, 9 and the December 2012 release of the 1000 Genomes European reference panel. After imputation, approximately 16 M SNPs passed post-quality control measures (HWE>1e-6, Missing <10%). It was from these genotyped and pruned SNPs that genotypic information was derived.

Selection of GWAS SNPs

BMI risk SNPs were obtained from The Genetic Investigation of ANthropometric Traits (GIANT) consortium’s association analysis, which revealed 18 new BMI-increasing loci, in addition to 4 prior known waist and height loci, and 10 previously identified BMI loci.10 Additional SNPs were provided by GIANT’s 2013 genome-wide meta-analysis that identified 11 new anthropometric loci, of which those attributing significantly to the overweight, obesity class I, and obesity class II categories were used in this analysis.11 Individual loci used to construct the GRS are listed in Supplementary Table S1.

Genetic risk score

After genotyping, imputation, and pruning, 35 of the 39 candidate SNPs were available for the construction of a GRSBMI. SNPs rs4771122 (hg38 chr13:27446043:G>A), rs4735692 (hg38 chr8:75703428:A>G), rs3810291 (hg38 chr19:47065746:G>A), and rs2287019 (hg38 chr19:45698914:C>T), tagging MTIF3, HNF4G, TMEM160(Q), and QPCTL respectively, failed quality control procedures in all five sub-cohorts, and were thus excluded from analysis. If a locus failed pruning in a particular sub-cohort, it was coded as missing (NA) and thus excluded. The GRSBMI was an unweighted sum of predisposing alleles. The maximum number of risk alleles possessed by one individual was 32 and the minimum was 3. The average number of risk alleles possessed by any one individual was 15 (mean±standard deviation (SD)=14.99±3.44). The BMI-increasing loci were assumed to have an additive effect, and the GRS was normally distributed.

Statistical analysis

Each of the 35 post-quality control loci were coded as 0, 1, or 2, according to the number of effect alleles present. The GRSBMI was computed by the sum of the number of BMI-increasing alleles for each subject. Traditional CAD risk factors (TRFs) used in this analysis included sex, current smoking status, plasma triglycerides (TG), LDL cholesterol (LDLc), and HDL cholesterol (HDLc). As noted above, patients with diabetes mellitus were not recruited. Owing to study design and intentionally very different ages of cases versus controls, we could not use age as a covariate in the analysis. Individuals were stratified by risk level: those with more than 15 alleles (mean for the entire population) were denoted as high risk (HR), while those with 15 or fewer risk alleles were denoted as low risk (LR). Those at or above the 90th percentile (≥20 risk alleles, n=951) were denoted as very HR (VHR) and those at or below the 10th percentile (≤10 risk alleles, n=893) were denoted as very LR (VLR). Logistic regression models were used to examine associations with CAD. Univariate models consisted of CAD as a binary response variable and GRSBMI as a continuous predictor, while multivariate models consisted of CAD as a binary response variable and GRSBMI, sex, smoking status, TG, LDLc, and HDLc as predictors. The explained variance of models was obtained by computing Nagelkerke’s pseudo-R2 for logistic regressions (http://CRAN.R-project.org/package=fmsb). Logistic regression models were also used to analyze the differences between HR and LR, in which case HR was coded as 1 and LR was coded as zero, as were VHR and VLR. Receiver operator characteristics (ROC) curves were generated by coding cases as 1 and controls as 0. Sensitivity, specificity, and accuracy were calculated from the resultant 2 × 2 table. The area under the curve (AUC) was used as a measure of predictive accuracy for the classifier. Non-parametric methods developed by Delong et al.12 were used to test for significant differences between ROC AUCs. ROC analysis was performed in R, using the PredictABEL13 and pROC14 packages. Additionally, reclassification tables were constructed using various classifiers in PredictABEL in R. All analyses were performed in PLINK: Whole genome Association Analysis Toolset,15 and R version 3.1.0 – ‘Spring Dance’ (http://www.R-project.org/).

Genetic correlations between obesity and CAD

GWAS data can be used to estimate the phenotypic variance explained by SNPs by comparing a matrix of pairwise genomic similarity to a matrix of pairwise phenotypic similarity in unrelated individuals using a random-effects mixed linear model. A bivariate extension of this method can be used to estimate the genetic correlation between two traits (rg) by relating the pairwise genetic similarity matrix to a phenotypic covariance matrix between traits 1 and 2. We used the bivariate REML analysis implemented in GCTA to estimate the genetic correlation between obesity and CAD. Any pair of individuals whose genetic similarity is equal to or greater than a fourth cousin was removed (pairwise relatedness >0.025) prior to the analysis, and first and second ancestry principal components were used to correct for population stratification and for genotyping artifacts.

Bivariate analysis was performed in OHGS_A2, OHGS_B2, OHGS_C2, and CCGB_2 cohort, and the genetic correlations were combined and weighted mean of genetic correlation was calculated based on fixed-effect model as follow:

Where wi is the inverse variance of the ith study and n is the number of studies

Standard error (SE) of combined effect is:

and P-value was calculated as:

where Ф is the standard normal cumulative distribution function.

Data are available at http://www.gwascentral.org/study/HGVST1831

Results

BMI is strongly associated with CAD case–control status

General population characteristics by case–control status are provided in Table 1. The prevalence of obesity (BMI>30 kg/m2) (χ2=263.15, P<2.2 × 10−16) was much greater in CAD cases versus controls. The control subjects were older than cases (owing to the original design of the CAD GWAS that sought to recruit young CAD cases and older healthy control subjects) and had lower plasma TGs and higher HDLc. Cases also had a higher smoking prevalence versus controls (χ2=277.19, P<2.2 × 10−16).

Table 1 Characteristics of coronary artery disease cases and healthy control subjects
Full size table

Genetic risk for obesity is associated with CAD

In the whole population, the constructed GRSBMI significantly (P=4.33 × 10−10) associated with CAD in a univariate model, with a per allele odds ratio (OR) of 1.04 (95% CI=1.026–1.051). The constructed GRSBMI also positively and significantly (P=2.12 × 10−12) associated with CAD status in multivariate models adjusted for TRFs, with a per-allele OR of 1.06 (95% CI 1.042–1.076) (Table 2). In the univariate model alone, GRSBMI explained ~0.5% of variation in CAD (Nagelkerke’s R2=5.490 × 10−3), while the multivariate model including TRFs explained ~27% of variation (Nagelkerke’s R2=0.2701). When GRSBMI was omitted from the multivariate model, ~26.25% of CAD variation was explained, a differential of ~0.75% which is similar to the univariate model. As an analogue, if BMI was included in the model and subsequently removed, we observed that BMI adds ~1.9% to the explained variance (Nagelkerke’s R2 with BMI=0.2817, Nagelkerke’s R2 without BMI=0.2625). Each risk allele resulted in an increased BMI of 0.0541 (95% CI=0.0202–0.0880), or ~156 g of weight for a 1.7 m (67 in) individual.

Table 2 Effect estimates and explained variance of models
Full size table

Genetic risk for obesity associates with CAD status after adjustment for measured BMI

To test whether GRSBMI contributed to CAD risk beyond measured BMI, separate models were constructed in which the score and BMI were either included as covariates or not. These parameters and values are provided in Table 2. When both GRSBMI and BMI were included as predictors, along with TRFs, the model explained 28.9% of CAD variation as compared with 26.25% of variation with TRFs alone. Thus, the GRSBMI and BMI explain 2.65% of CAD variation. As noted above, the addition of GRSBMI to traditional predictors explained 0.75%, and the addition of BMI explained ~1.9%, for a total of ~2.65%. Thus, GRSBMI provides information on CAD risk beyond BMI measurement per se.

Among those with a GRSBMI above versus below the average for the whole population, the mean number of risk alleles differed by 6.18 (95% CI=6.10–6.27). When considered as a binary trait, HR versus LR status significantly (P=2.64 × 10−9) predicted CAD, giving a univariate OR of 1.29 (95% CI=1.18–1.40) and a multivariate OR of 1.49 (95% CI=1.33–1.66, P=8.65 × 10−13). Of the CAD cases, 45.6% were HR (>15 risk alleles) as compared with 39.4% of controls (P=2.9 × 10−9). When subjects were then stratified by the top and bottom deciles for GRSBMI (VHR and VLR), as expected, the differences were more pronounced. In a univariate model, VHR versus VLR status was associated with an OR for CAD of 1.65 (95% CI=1.36–1.99, P=2.31 × 10−7), while in a multivariate model, VHR versus VLR risk status was associated with an OR of 1.98 (95% CI=1.52–2.59, P=4.55 × 10−7). Again a higher proportion of cases were in the upper decile of GRSBMI as compared with controls (χ2 P=2.75 × 10−7) (Table 3). When SNPs were analyzed individually, none survived multiple testing correction for association with CAD.

Table 3 Association of upper versus lower deciles for BMIGRS with CAD
Full size table

Effect of addition of GRSBMI to traditional CAD risk factors

TRFs yielded an AUC of the ROC of 0.7642 (95% CI=0.7530–0.7754), predicting CAD moderately well. The addition of GRSBMI to TRFs provided a 0.35% increase in predictive accuracy (AUC=0.7677, 95% CI=0.7566–0.7788). This difference was small, but significant (DeLong’s test for correlated ROC curves P=3.38 × 10−3). Similarly, the addition of BMI to TRFs resulted in a 0.9% increase in predictive accuracy (AUC=0.7731, 95% CI=0.7621–0.7841), a small but significant (DeLong’s test for correlated ROC curves P=1.08 × 10−8)12 addition to the model’s predictive accuracy. Together, GRSBMI and BMI added ~1.2% predictive accuracy to the model (AUC=0.7761, 95% CI=0.7652–0.7871), a significant improvement (P=1.12 × 10−8).

The addition of GRSBMI to TRFs provided a continuous net reclassification index (NRI) of 16.54 % (95% CI=11.82–21.26%, P<0.0001) (Supplementary Table S2). It is important to note that the reclassification tables constructed do not follow clinically relevant strata of 5, 10, and 20 percent 10-year risk, as our models did not involve Cox-proportional hazard modeling with a 10-year longitudinal analysis. Thus, the NRI although relevant, does not have direct clinical application because the prevalence of CAD in the study population is much higher than in a normal population, and thus, frequencies and risks are highly inflated. The addition of GRSBMI to TRFs also provided an integrated discrimination improvement of 0.0058 (P<0.0001). Full reclassification results are provided in Table 4.

Table 4 Reclassification table comparing predicted CAD risk with and without GRSBMI
Full size table

Genetic correlation between BMI and CAD

Finally, we estimated the genetic correlation (rg) between BMI and CAD by relating the pairwise genetic similarity matrix to a phenotypic covariance matrix between the traits 1 and 2, using the bivariate REML analysis implemented in GCTA. These analyses provide strong evidence of genome-wide pleiotropy between obesity and CAD (Table 5).

Table 5 Genetic correlations between obesity and CAD
Full size table

These data are available at http://www.gwascentral.org/study/HGVST1831

Discussion

Here, we have performed a large Mendelian randomization study to assess the contribution of genetic risk for obesity to CAD in a population without diabetes mellitus and when controlling for other TRFs highly associated with obesity including plasma TGs and HDLc.16 We report that a GRSBMI based on 35 BMI risk alleles identified by the GIANT consortium significantly (P=2.12 × 10−12) predicts CAD; each allele predisposing to higher BMI yields an OR for CAD of 1.06 or a 6% increase in risk. Each BMI risk allele predisposes to a mean 156 g increase in body weight for a 1.7 m individual. Subjects with a GRSBMI above 15 were more likely (OR=1.49) to be CAD cases as compared with those in the lower half of the GRSBMI distribution. Those in the upper versus lower decile for GRSBMI had twice the prevalence of CAD (OR=1.98). This finding confirms the importance of genetic contributions to both obesity and its vascular sequelae.

In the current study, no individual BMI risk variant significantly associated with CAD, likely owing to inadequate statistical power. Here, a GRSBMI explained ~0.75% variation for CAD when BMI was included in the model, increasing the explained variance from 26.25% for TRFs alone, 28.16% with TRFs and BMI, to 28.86% with TRFs, BMI, and GRSBMI. This indicates that the BMI GRS is predictive over and above measured BMI as discussed below. Our assumption here is that the BMI risk SNPs associate with obesity over the course of life and that the CAD risk due to obesity is cumulative over a lifespan. Similarly, variants associated with LDLc concentrations in the PCSK917or NPC1L118 genes are more strongly associated with CAD than are associated differences in measured LDLc. For any risk factor, including cigarette smoking, LDLc, or obesity, life years of exposure can be considered more informative than a single measure.

This study was carefully designed, included an extended number of BMI risk alleles, and corrected for important BMI-associated CAD risk factors. In Mendelian randomization studies, bias by confounding is minimized, as the distribution of genetic variants are, in the general case, independent from the behavioral and environmental factors that would usually confound associations between risk factors and disease phenotypes. This approach also allows causal relationships to be established. However, a number of caveats must be acknowledged. For example, possible pleiotropic effects of BMI risk variants may be directly linked to the pathogenesis of CAD. FTO has been studied in this regard and direct effects on atherogenesis have been hypothesized2, 3, 4, 5 but not well substantiated. It is, however, conceivable that the BMI risk alleles are in weak indirect pleiotropy with any of a number of phenotypes which could affect CAD risk or that a fraction is in direct pleiotropy. It is beyond the scope of this study to explore such causal pathways. It appears more likely that the additional variation in CAD explained by the GRSBMI versus BMI alone is reflective of lifetime risk and adiposity over a lifetime may differ substantially from BMI measured at one discrete point in time. The use of a GRS as a proxy measure for lifetime BMI can avoid this dilution bias, and thus the variance observed may be a result of life course BMI rather than an incidental measure. This relationship to lifetime risk has been used to explain the observations that coding variants in the PCSK9 gene contribute to alterations in CAD risk that are greater than that predicted by the associated differences in LDLc concentrations.17

Although we excluded individuals with diabetes mellitus and adjusted for TRFs, this adjustment was based on single measures for each. Beyond effects on conventional risk factors,16 GRSBMI-associated risk may also derive from additional processes such as pro-inflammatory milieu of adipose tissue19 or differences in the gut flora.20 Whether through pleiotropy, effects on lifetime adiposity measures, or factors not yet considered, this study provides strong evidence that cumulative risk for obesity associates with CAD. The addition of the GRSBMI to TRFs increased predictive accuracy by approximately 0.35% a small but significant (P=3.38 × 10−3) difference. This is apparent even when BMI is included added to TRFs, again showing that the GRSBMI predicts BMI, which itself adds around 0.9% predictive accuracy. Together GRSBMI and BMI yielded a continuous NRI of 16.5% (P<0.0001) and an integrated discrimination improvement of 0.0058 (P<0.0001), a significant improvement in the model.

For this study, we recruited elderly asymptomatic control subjects, because younger subjects may harbor a significant burden of occult atherosclerosis. To the knowledge of the authors, none of the alleles used in this analysis have been associated with longevity, and thus, differing allele frequencies between cases and controls should not contribute significantly to the results of this study. It should be noted, however, that in elderly subjects, TRFs may be less strongly associated with CAD,21 and the observed effect of the GRSBMI may differ from those encountered in a prospective cohort study.

In summary, in a Mendelian randomization study, we demonstrate a causal relationship between BMI and CAD risk. We demonstrate that an individual with an above average genetic risk for obesity has a 50% greater risk for CAD, while those individuals in the top GRSBMI decile have twice the CAD risk of those in the bottom decile. We additionally show that a GRSBMI has predictive value in addition to BMI per se. The addition of GRSBMI to TRFs provided a small but significant increase in predictive accuracy, NRI, and integrated discrimination improvement. Finally, we provide evidence of genome-wide pleiotropy between obesity and CAD. These analyses further strengthen a body of evidence that obesity is causal risk factor for CAD.