Partitioning the Pleiotropy Between Coronary Artery Disease and Body Mass Index Reveals the Importance of Low Frequency Variants and Central Nervous System–Specific Functional Elements
Background: The objective of this study is to investigate the extent and nature of pleiotropy between coronary artery disease (CAD) and body mass index (BMI).
Methods: We examined the contribution of genome-wide single-nucleotide polymorphisms (minor allele frequency ≥0.01) to co-occurrence of CAD and BMI in a sample of genetically unrelated 8041 subjects (genetic resemblance ≤0.025) of European ancestry using mixed-linear-models. We further partitioned the estimated pleiotropy according to biological features to gain insight into the nature of pleiotropy between CAD and BMI.
Results: We found significant (P<0.0001) positive genetic correlation between CAD and BMI (rg=0.60). The estimated pleiotropy explained 68% of phenotypic correlation, and it was not proportionally distributed across the chromosomes; notably, chromosome 10 contributed more; whereas, chromosomes 11 and 14 contributed less to pleiotropy than expected given their chromosomal length. We noted that a large proportion (63%; P=0.002) of the pleiotropy is attributed to single-nucleotide polymorphisms with low allele frequency (minor allele frequency <0.05). Of note, pleiotropy was enriched among central nervous system genes and genes of metabolic pathways. Further analyses revealed that these effects are more pronounced in the proopiomelanocortin pathway and genes involved in carbohydrate metabolism. After genome-wide association study meta-analysis, only single-nucleotide polymorphisms downstream of the MC4R gene were found concordantly associated with (P<5×10–8) BMI and CAD with lead single-nucleotide polymorphism being rs663129 (combined P=2.7×10–65). Finally, partitioning the pleiotropy according to functional elements pointed to the importance of superenhancers and notably brain-specific superenhancers.
Conclusions: Genome-wide pleiotropy substantially contributes to co-occurrence of CAD and obesity, and it is highly enriched among low frequency variants and central nervous system–specific functional elements.
- body mass index
- central nervous system
- coronary artery disease
- gene frequency
- polymorphism, single nucleotide
In this study, we investigated the contribution of genome-wide single-nucleotide polymorphisms to co-occurrence of coronary artery disease (CAD) and body mass index in a sample of 8041 subjects. We found that the co-occurrence of CAD and body mass index is largely because of many shared genetic factors (pleiotropic loci) that are scattered throughout the genome. We also investigated the nature of pleiotropy between body mass index and CAD in a variety of ways. We found that the pleiotropy between body mass index and CAD can be largely attributed to single-nucleotide polymorphisms with low allele frequency. Pleiotropy is enriched among central nervous system genes and genes of metabolic pathways, most notably the proopiomelanocortin pathway and genes involved in carbohydrate metabolism. Finally, we found that variation in gene expression is an important mechanism underlying susceptibility to obesity and CAD; in this regard, our pleiotropy partitioning analysis pointed to the importance of superenhancers and notably brain-specific superenhancers.
Obesity is a heritable, modifiable, risk factor for coronary artery disease (CAD). Clustering of CAD and body mass index (BMI) in adults has been observed in epidemiological studies and clinical settings.1 Whether this co-occurrence results from the pleiotropic effects of genetic factors has proven a complex question. Nonetheless several shared mechanisms that are important in the pathogenesis of these traits have been reported,2 and Mendelian randomization studies have shown that the increased genetic risk of obesity has a predictive value for CAD.3,4
With the availability of genome-wide genotype data from genome-wide association study (GWAS) studies, it is now possible to calculate the genetic resemblance among individuals and estimate its contribution to co-occurrence of 2 traits using mixed-linear-models (MLMs).5–7 This approach quantifies the overall contribution of single-nucleotide polymorphisms (SNPs) to (co)variation (ie, variation and covariation) of traits without testing the SNPs individually. This is an important feature because in a typical GWAS, hundreds of thousands or millions of SNPs are tested separately, and to avoid false-positive discoveries, the stringent Pvalue threshold of 5×10−8 is used to report a significant finding. Therefore, if there are many loci each with a small effect (as is the typical feature of complex diseases), most of these genetic variants will fail to reach GWAS significance.8 Moreover, MLMs allow to estimate the proportion of pleiotropy that is explained by a subset of SNPs that is defined according to functional annotation, pathways, allele frequency, etc, and therefore provide the opportunity to leverage annotation information and genome-wide genotype data to gain insight into the nature of pleiotropy between traits.6,7
In this study, exploiting the 1000 Genomes imputed data, we used the MLM approach to investigate the extent of genome-wide pleiotropy between BMI and CAD in a sample of 8041 unrelated subjects from a genetic point of view (genetic resemblance ≤0.025). Next, we partitioned the estimated pleiotropy by chromosome, minor allele frequency (MAF), modules, and SNP annotation to gain insight into the nature of pleiotropy between BMI and CAD.
The complete description of Methods is available in the Data Supplement. The data, analytic methods, and study materials are available on request for purposes of reproducing the results or replicating the procedure. This study was performed in a sample of subjects of European ancestry that were initially assembled by the Ottawa Heart Institute as part of a large study on the genetics of CAD.9,10 Cases were <55 years (men) and <65 years (women) at the time of myocardial infarction, coronary artery bypass graft, percutanaeous intervention, or angiography indicating a stenosis of ≥50% in at least 1 epicardial vessel. Subjects were excluded if they had severe pulmonary hypertension or congenital heart disease or were diabetic. Controls were elderly subjects who either had a computed tomographic angiography or angiogram demonstrating no stenosis of >50% or were asymptomatic for cardiovascular disease. Our study was approved by the Research Ethics Board of the Ottawa Heart Institute, and all participants provided written informed consent.
SNPs with missingness >1%, Hardy–Weinberg equilibrium <0.0001, MAF <0.01, and subjects with >1% genotype missingness or discrepancies between the reported sex and sex determined from the X chromosome were excluded. Genotypes were prephased and followed by genotype imputation based on the 1000 Genomes reference panel (phase I; release 3). After genotype imputation, we excluded SNPs with MAF <1%, missing values in >1% of individuals, or Hardy–Weinberg equilibrium P <0.0001 and generated genomic relatedness matrix (GRM) from the genotype data using GCTA software (v1.25).7 Next, we performed identity-by-state (IBS) pruning by excluding one of each pair of individuals with an estimated, genome-wide IBS (genetic resemblance) >0.025 and retained a subset of 8041 unrelated individuals. Finally, we generated a GRM using all SNPs (3 163 082 SNPs) and estimated its contribution (genetic correlation) to co-occurrence of BMI and CAD using genomic relatedness based restricted maximum likelihood (GREML) approach implemented in GCTA. Genetic correlation (rg) is defined as:
where Vg1g2 is the additive genetic covariance between the 2 traits, and Vg1 and Vg2 are the additive genetic variances of trait 1 and trait 2 as estimated by GREML. The SE of rg was then computed from the estimated errors of Vg1g2, Vg1, and Vg2 using delta method implemented in GCTA. Furthermore, we partitioned the estimated pleiotropy by chromosome, MAF, modules, and SNP annotation to gain insight into the nature of pleiotropy between BMI and CAD.
Heritability and Correlation Estimates
Our data set consisted of a total of 8041 unrelated subjects (genome-wide IBS ≤0.025) with available postimputed autosomal genotype data for 3 163 082 genome-wide SNPs with MAF ≥0.01, Hardy–Weinberg equilibrium ≥0.0001, missingness <0.01, and info >0.4. In this data set, we have 8025 individuals with BMI data, 7619 individuals with CAD data, and 7603 individuals with nonmissing phenotypes for both BMI and CAD. General characteristics of the subjects are shown in Table 1. IBS pruning did not change the distribution of our data. This is expected considering that the GCTA algorithm randomly excludes 1 member of each pair of individuals with genome-wide IBS >0.025. The prevalence of CAD was 70% in men and 43% in women (P<0.001; Table 1) in the final sample. As expected, CAD cases had higher BMI than controls (P<0.001; Table 1); controls were older than cases according to study protocol because controls were selected from asymptomatic elderly subjects (P<0.001; Table 1).
We conducted principal component analysis of ancestry and mapped our sample with African, European, South Asians, and East Asians population samples from 1000 Genomes Project. As it is depicted in the principal component analysis graph of ancestry (Figure I in the Data Supplement), our subjects clustered within the European sample indicating all subjects are of European ancestry. Next, we used the GREML method implemented in MTG2 program5 to compute the genetic (co)variance (genetic variance and genetic covariance) explained by genome-wide SNPs. Our study was adequately powered (power ≥80%) to detect SNP heritabilities as small as 0.11 and genetic correlation as small as 0.4 (Figure 1) at α (significance level)=0.05.
We used the notation hg2 to depict the proportion of variance in BMI and liability to CAD explained by common genome-wide SNPs. hg2 represents the lower bound of narrow sense heritability because it only represents variation because of common SNPs. We also wish to clarify, in this study, the term heritability means hg2 (SNP-based heritability). After fitting the GRM generated based on all 3 163 082 SNPs in the MLM model, we found that 12% of BMI variation (Table 2, A) and 22% of variance in liability to CAD are explained by all SNPs (Table 2, B).
Consistent with the positive phenotypic association between CAD and BMI (Table 1), we also found significant positive genetic correlation (rg=0.6; SE=0.13; Table 2, C), which means subjects with higher BMI show higher genetic similarity in the CAD case group as compared with the controls; whereas, subjects with lower BMI show higher genetic similarity in CAD control group as compared with the cases (P=5×10–6; Table 2, C). Of note, the estimated genetic covariance on heritability scale which is known as coheritability was 0.1 (SE=0.02); coheritability allows comparison of the shared liability attributable to SNPs on the same scale as heritability (hg2) and is the indicator of pleiotropy between 2 traits.6 Using equation 8 (Data Supplement), next we calculated the phenotypic correlation between BMI and CAD (rp=0.14) and found that 68% of phenotypic correlation is attributed to pleiotropy (coheritability).
Both BMI and CAD showed strong sexual dimorphism (P<0.0001); namely, the mean BMI was higher in men and the prevalence of CAD was higher in men as compared with women. We thus investigated whether these sexual dimorphisms have genetic underpinnings by splitting the samples into male and female groups. Next, we did bivariate GREML analysis to measure the magnitude of genetic correlation between the 2 groups for CAD and BMI. Under this condition, the 2 independent subsets are related through the coefficients of similarity calculated from SNPs.6 The genetic correlation (rg) explained by SNPs between the sexes was 1 (SE=0.5) for BMI and 0.9 (SE=0.3) for CAD, not significantly different from 1 (P>0.3; Figure II in the Data Supplement). This indicates that the observed sexual dimorphisms for BMI and CAD are not because of common genomic variants.
Partitioning of Genetic (Co)variance by Chromosome
To investigate the contribution of each chromosome to pleiotropy and heritabilities of BMI and CAD, we generated the GRMs based on the SNPs on each chromosome and partitioned the total genetic (co)variance onto individual chromosomes by fitting the GRMs of all the chromosomes simultaneously through a joint model. The scatter plot of partitioned heritabilities (h2c) and genetic covariance by chromosomes are shown in Figure 2. We found that chromosomal length explains a small proportion of variation in h2c for BMI (coefficient of determination or R2=10%; P=0.09) and CAD (R2=20%; P=0.02), as well as the genetic covariance (R2=20%; P=0.03). In addition, the correlations of h2c of BMI, h2c of CAD, and their genetic covariance with chromosomal length were significantly different (P<0.003) from 1 (H0: r=1), which indicates that the genetic factors underlying these traits are not proportionally spread across the chromosomes. Notably, chromosome 10 contributed more to the genetic covariance of BMI and CAD than expected given its length, whereas chromosomes 11 and 14 contributed less (P<0.05; Figure 2C; Table I in the Data Supplement).
(Co)heritability Partitioning by SNP Types
A total of 43% of variants in our data set were genic, residing in exonic, intronic, noncoding transcripts or regulatory regions near the gene (Figure 3A). Genic variants accounted for 79% of BMI heritability and 61% of CAD heritability which was notably higher than intergenic variants (P<0.007; Figure 3A; Table II in the Data Supplement). Genic variants also explained a large proportion of coheritability (61%); however, given the large SEs, the difference was not significant as compared with intergenic SNPs (Figure 3A; Table II in the Data Supplement).
Next, we asked whether the (co)heritabilities (heritability of BMI and CAD as well as coheritability between them) are proportionally spread across the allele frequency spectrum? For this purpose, we allocated SNPs into 5 MAF bins and estimated their joint contribution to (co)heritabilities. We found that a significant proportion of coheritability is explained by low allele frequency SNPs (0.01≤MAF<0.05; Figure 3B; Table III in the Data Supplement). Low allele frequency SNPs represented 25% of SNPs but explained 63% of the coheritability (SE=13%; P=0.002). However, as shown in Figure 3B, BMI and CAD heritabilities explained by each MAF bin were within the expected range. This suggests that the genetic determinants of BMI and CAD are proportionally spread across the allele frequency spectrum.
(Co)heritability Partitioning by Gene Sets
We quantified the contribution of several gene sets previously implicated in the pathogenesis of these traits.2,11 These included inflammatory response genes reported by Calvano et al,12 genes involved in metabolic processes as depicted in the Kyoto Encyclopedia of Genes and Genomes database, and central nervous system (CNS) genes identified by Raychaudhuri et al.13 We also investigated the contribution of known CAD and obesity GWAS loci (P<5×10–8) from the GWAS Catalog.
Consistent with the notion that obesity and CAD are inflammatory conditions,11,14 genes involved in inflammation explained a significant (P≤0.01) amount of heritability of BMI and CAD. This indicates that inflammation dose contributes to the development of obesity and CAD and is not merely a consequence of these phenotypes. Inflammatory response genes also explained 14% of coheritability, which is higher than the proportion of SNPs they represented (1.1%); however, given the large SE (8.3%), the difference was not statistically significant (Figure 4A; Table IV in the Data Supplement).
Known CAD loci from the GWAS catalog explained a considerable amount of CAD heritability (16%; P=3e−5), but their contribution to heritability of BMI and their coheritability were insignificant. Similarly, obesity GWAS loci explained a significant proportion of heritability of BMI (18%; P=0.01), but their contribution to genome-wide heritability of CAD and the coheritability between them was not notable (Figure 4A; Table IV in the Data Supplement).
Given that number of centrally mediated risk factors,15–17 such as depression and substance use, have been reported for CAD and obesity, next we investigated the contribution of CNS genes to (co)heritability of CAD and BMI. CNS genes explained 31% of coheritability that was significant given the proportion of SNPs represented by them (P=0.03; Figure 4A; Table IV in the Data Supplement). Of note, we also did a fixed-effect meta-analysis using CAD and BMI GWAS results. Here, we used the summary statistics from the 1000 Genomes–based GWAS meta-analysis of CAD9 and BMI.18 We found that only SNPs in a region downstream of the CNS gene, MC4R, concordantly associated with BMI and CAD (P<5×10–8) and lacked heterogeneity (I2=0) in our meta-analysis (Figure 5; Table V in the Data Supplement). The lead SNP in this region, rs663129, surpassed GWAS significance (P=2.7×10–65; allele=A; β=0.06; and SE=0.003) in our meta-analysis; allele A of this SNP increases risk of both CAD (β=0.058; P=3.2×10–8) and obesity (β=0.055; P=3×10–57). Because MC4R functions through the proopiomelanocortin pathway, we examined the contribution of genes in the proopiomelanocortin pathway (as depicted in the PANTHER Pathways database [http://www.pantherdb.org]) and found that this pathway significantly (P=0.01) contributes to the observed pleiotropic effect between CAD and BMI (Figure 4B; Table VI in the Data Supplement)
Disturbances in metabolic processes, such as dyslipidemia and glucose dysmetabolism, are known to influence obesity and CAD2; therefore, we investigated the contribution of metabolic pathways genes to (co)heritability of BMI and CAD. Genes involved in metabolic pathways explained 11% of pleiotropy that was significant (P=0.03) given the proportion of SNPs represented by them (Figure 4A; Table IV in the Data Supplement). Among these genes, we identified a subset of 28 genes that positively contributed to the observed pleiotropic effect (Figure 4B; Table VI in the Data Supplement). We found significant enrichment of protein–protein interaction among these genes (clustering coefficient=0.79; protein-protein interaction enrichment P=8.6e−10; Figure III in the Data Supplement). Go-Term enrichment analysis based on biological processes indicated that the identified genes are mainly involved in carbohydrate metabolism (Figure III in the Data Supplement).
Pleiotropy Partitioning by Functional Elements
Next, we investigated the nature of pleiotropy between BMI and CAD by examining a broad set of functional elements. Here, we examined functional categories from Finucane et al19 and estimated their contributions to the pleiotropy of BMI and CAD (Table VII in the Data Supplement). In summary, these categories include coding, untranslated regions, promoter, intronic and intergenic regions; the histone marks (H3K4me3, H3K4me1, H3K9ac, and H3K27ac); DNase I hypersensitivity sites; combined chromHMM and Segway predictions of chromatin elements from Encyclopedia of DNA Elements data; conserved regions in mammals; superenhancers; and enhancers with balanced bidirectional capped transcripts identified in the FANTOM5 panel of samples. To prevent the estimates from being biased upward by enrichment in nearby regions, 500-bp windows around each of the annotations were also included; moreover, for each DNase I hypersensitivity sites, H3K4me3, H3K4me1, and H3K9ac site, a 100-bp window around the chromatin immunoprecipitation-seq peak was added as an additional category.
The proportion of SNPs and coheritability attributed to each annotation category is presented in Table VII in the Data Supplement. We found large enrichment of coheritability in superenhancer regions with 14% of SNPs explaining 59% of coheritability (95% CI: 0.32–0.86; Table VII in the Data Supplement). Superenhancers are cluster of transcriptional enhancers that drive cell-type–specific gene expression programs and therefore can be examined to track the key cells contributing to pleiotropy of BMI and CAD. For this purpose, we obtained the genomic coordinates of superenhancers across a variety of human tissue/cell types from dbSUPER database20 and investigated the contribution of each cell-specific superenhancer category to the coheritability of CAD and BMI. The proportion of SNPs and coheritability attributed to each category is presented in Table VIII in the Data Supplement. The categories that passed the correction for multiple testing (false discovery rate-corrected P<0.05) were all brain-specific tissues, including hippocampus middle, inferior temporal lobe, and cingulated gyrus (Figure 6).
We also merged the superenhancer coordinates from all cells and investigated its contribution to coheritability of BMI and CAD as the combined category. Coheritability was significantly enriched in the combined category, with 15% of SNPs explaining 61% of coheritability (SE=0.14; P=0.001; Table VIII in the Data Supplement) which further verifies our initial finding presented in Table VII in the Data Supplement. Finally, 7527 nonsynonymous SNPs present in our data set explained 6% of coheritability (95% CI: 0%–24%) that was within the expected range given the proportion of SNPs they represent (0.23%).
Traditionally, genetic (co)variances (genetic variances and genetic covariance) of traits are estimated from phenotypic similarity between relatives. However, relatives often share both genes and a similar environment and so it is difficult to completely separate the genetic (co)variance from the environment. Furthermore, achieving genetically informative samples of sufficient size is generally more difficult in family-based designs as compared with a population.6,21 Recently, Lee et al22 developed the bivariate GREML approach to obtain unbiased estimates of the genetic correlation between pairs of traits using population-based studies with genome-wide SNP data. The method has been previously validated in simulation studies22,23 and used in several studies, such as to estimate the genome-wide pleiotopy between various diseases from Wellcome Trust Case Control Consortium,22 to dissect the shared genetic architecture between psychiatric disorders,6 to study between-sex genetic heterogeneity for human height and BMI,15 or to investigate the genetic correlation between schizophrenia and height.23
In this study, leveraging phased haplotypes from the 1000 Genomes Project, we used the bivariate GREML approach to investigate the contribution of genome-wide SNPs to (co)heritability (heritability and coheritability) of BMI and CAD in a sample of 8041 genetically unrelated subjects of European ancestry. We found that genome-wide SNPs (MAF ≥0.01) explain 22% of CAD risk and 12% of BMI variance. In addition, the estimate of SNP-based genetic correlation (rg) which reflects genome-wide pleiotropy tagged by common SNPs was large between BMI and CAD. In this regard, after analyzing publicly available GWAS summary statistics for BMI and CAD, Bulik-Sullivan et al24 also reported a significant positive genetic correlation between BMI and CAD. Besides, using a similar approach, Gusev et al25 reported hg2=0.25 for CAD in Wellcome Trust Case Control Consortium, and Vattikuti et al26 reported hg2=0.14 for BMI in ARIC study (The Atherosclerosis Risk in Communities Study that investigates the pathogenesis of atherosclerosis in individuals aged 45–64 years). Of note, our heritability estimate for BMI is much lower than estimates from GIANT18 (The Genetic Investigation of ANthropometric Traits) and deCODE studies27 (hg2≈0.22) for BMI (using samples drawn from general population), and this is because the cases and controls in our cohorts were initially selected for the purpose of studying the genetics of CAD and therefore, the BMI phenotype from these subjects is a subset and not a representative of the whole spectrum of BMI as seen in general population.
It is unlikely that population stratification or cryptic relatedness caused the observed effects as we removed subjects with genetic resemblance <0.025 and kept only subjects of European ancestry. We assume that the observed effects are not because of confounding environmental factors either; this is because the genotype of a person for a trait is randomly determined through the meiosis (Mendelian randomization). In addition, the bivariate GREML approach computes the genetic parameters after decomposing the phenotypic (co)variance into its genetic and residual (nongenetic) components; furthermore, unrelated subjects are less likely to share a similar environment as compared with pedigree and family-based designs.
We did not detect significant differences in heritability of CAD and BMI between men and women, which indicates that the majority of common variants that contribute to CAD risk or BMI variance are shared between sexes. This is in agreement with previous GWAS studies that did not identify variants with sex-different effects for these traits.9,28 In addition, we found that genetic factors underlying (co)heritability of BMI and CAD are not uniformly spread across the genome proportional to chromosomal lengths. In this regard, a previous study also found that chromosomal length explain small proportion (R2=0.08) of variation in h2c (heritability by chromosome) for BMI.7 However, this phenomenon has not been always observed, for example, genetic factors underlying human height and QT interval are reported to be proportionally distributed across the chromosomes.7 The observation that chromosome 10 contributed more to coheritability than expected provides hint for future bivariate association studies that aim to detect pleiotropic SNPs underlying CAD and BMI; of note, we found a significant enrichment of coheritability among superenhancers located on chromosome 10.
Evidence from sequencing studies and selection theories indicates that genetic variants with high penetrance that are at the lower end of allelic spectrum are important to the pathogenesis of complex diseases.29 In this regard, we found that the pleiotropy between BMI and CAD is mainly attributable to low allele frequency SNPs (MAF <0.05). Because it is difficult to call lower frequency SNPs through the genotype imputation as compared with common SNPs, therefore, the design of genotyping/sequencing panels focused on lower frequency SNPs could accelerate identification of causal variants that contribute to pleiotropy of BMI and CAD.
After the meta-analysis of publically available CAD and BMI GWAS results, we only found SNPs downstream of the CNS gene, MC4R, that were concordantly associated with both traits. It is reported that SNPs in this region influence the expression of MC4R possibly through the methylation of MC4R promoter.30 Although MC4R is a well-replicated locus for BMI, dozens of other loci have also shown genome-wide association for BMI, including the renowned FTO gene.18 However, no SNPs at the FTO locus reached genome-wide significance for CAD (Figure IV in the Data Supplement). One explanation could be that the mechanisms whereby MC4R SNPs contribute to obesity also increase the liability to CAD, whereas the mechanisms by which FTO SNPs contribute to obesity do not affect the CAD risk. Of note, it has been shown that the lead SNPs within the FTO gene contribute to obesity by regulating the expression of IRX3/IRX5 genes in adipocyte tissues (independent of CNS),31 whereas, MC4R is highly expressed in the hypothalamus and has well-established roles in appetite and energy control through the proopiomelanocortin pathway.32 Here, pleiotropy partitioning also demonstrated that the CNS system significantly contributes to the pleiotropy, and this effect is more pronounced in the proopiomelanocortin pathway. We also found significant pleiotropy enrichment among genes involved in metabolic pathways, in particular carbohydrate metabolism that accords with the notion of impaired glucose metabolism and hyperinsulinemia as a mechanism linking obesity with cardiovascular disease.2 Of note, the proopiomelanocortin pathway and genes involved in metabolic pathways/carbohydrate metabolism could play interrelated roles undermining the observed pleiotropy. It has been shown that proopiomelanocortin pathway is involved in the regulation of glucose metabolism,33 and deletion of MC4R in mice leads to hyperinsulinemia before the onset of extreme obesity.34
Comprehensive analyses of known GWAS loci have pointed to the importance of regulatory regions of the genome.35,36 In this study, after analysis of the data from RegulomeDB (version 1.1),37 we also found a significant positive genetic correlation between CAD and BMI (rg=0.56; P=0.0002) in genomic regions involved in transcriptional activity (RegulomeDB score ≤5). In addition, after partitioning our (co)heritability estimates using the data from GenoCanyon (version 1.0.3)38 that predicts functional noncoding regions of genome at the nucleotide level, we noted that SNPs with higher functional scores contribute more to (co)heritability of CAD and BMI (fourth quartile or score >0.75; Figure 7; Table IX in the Data Supplement) whereas SNPs with lower functional scores show significant coheritability depletion (first quartile; Figure 7; Table IX in the Data Supplement). Several follow-up studies have identified specific functional elements that contribute disproportionately to the heritability of complex diseases.19,25 In our study, after examining a broad set of functional elements, we also found that superenhancer regions are highly enriched in genomic variants that contribute to the pleiotropy of BMI and CAD, and this effect is more pronounced in brain tissues, including hippocampus middle, cingulate gyrus, and inferior temporal lobe. These tissues are important in learning, memory, and emotion formation and processing and implicated in number of psychiatric disorders. Our finding is also important considering that numbers of centrally mediated traits,15–17 such as depression, and substance use are known to contribute to both obesity and CAD. Altogether, these evidences further add support to the notion that the CNS contributes to occurrence of cardiovascular disorders and associated risk factors.39,40
The recently proposed Omingenic model of inheritance states that variants in most of the genome contribute to heritability of a complex trait; however, contribution is higher in genomic regions that are active in relevant tissues.41 Under this model, pleiotropy is not a rare phenomenon, but it is a ubiquitous and inherent genetic feature of complex traits. Our findings are in line with predictions from this model. In addition, we found that pleiotropy between BMI and CAD is enriched in CNS-specific functional elements which adds support to the concept of network pleiotropy predicted by this model.
It is important to emphasize that MLM approach is a complement to the GWAS and not an alternative. We quantified the pleiotropy between CAD and BMI and showed that this pleiotropy is not proportionally distributed across the genome according to chromosomes, allele frequency, and genomic elements. An implication of this finding would be the design of custom genotyping/sequencing panels focused on those genomic features that showed significant pleiotropic enrichment and following it with bivariate association analysis to identify the causal SNPs. Such a joint procedure allows combining the merits of both approaches and also reduces the number of tests in Bonferroni correction. In addition, given the small effect sizes of most disease-associated SNPs and limited sample size of GWAS studies, our findings can also be used as prior information in variant prioritization analysis of CAD/BMI GWAS results.
BMI and the risk of CAD tend to increase as people age; however, given that our controls were older but leaner than cases (Table 1), and adjusting for age diminished known associations (eg, association of 9p21 locus and CAD), we did not include age as a covariate in our analyses. The occurrence of this pattern in our sample is because of the fact that controls were deliberately selected from asymptomatic elderly subjects. As such, it will be valuable if future studies explore our findings in samples randomly drawn from population.
In summary, in this study, we investigated the nature and extent of pleiotropy between obesity and CAD using >3 million genome-wide SNPs and MLM approach in a sample of 8041 genetically unrelated subjects of European ancestry. We found genome-wide pleiotropy substantially contributes to co-occurrence of CAD and obesity, and it is highly enriched among low frequency variants and CNS-specific functional elements.
We thank Dr Stanley Hazen, Director of the Cleveland Clinic GeneBank study, Drs Christopher Granger and Svati Shah of the Duke Cathgen study, and Dr Sonia Anand of the INTERHEART study for their contributions to this project.
Sources of Funding
This study is supported by Canadian Institutes for Health Research MOP-2390941 and OPB-134211 and Heart & Stroke Foundation of Canada T-7218 & BR-7519 to Dr McPherson.
The Data Supplement is available at http://circgenetics.ahajournals.org/lookup/suppl/doi:10.1161/CIRCGEN.117.002050/-/DC1.
Circ Genom Precis Med is available at http://circgenetics.ahajournals.org.
- Received July 18, 2017.
- Accepted January 16, 2018.
- © 2018 American Heart Association, Inc.
- Ahmad SI,
- Imam SK
- Tizón-Marcos H,
- Poirier P
- Cole CB,
- Nikpay M,
- Stewart AF,
- McPherson R
- Nikpay M,
- Stewart AFR,
- McPherson R
- Raychaudhuri S,
- Korn JM,
- McCarroll SA,
- Altshuler D,
- Sklar P,
- Purcell S,
- et al
- Gale CR,
- Batty GD,
- Osborn DP,
- Tynelius P,
- Rasmussen F
- Carney RM,
- Freedland KE
- Locke AE,
- Kahali B,
- Berndt SI,
- Justice AE,
- Pers TH,
- Day FR,
- et al
- Finucane HK,
- Bulik-Sullivan B,
- Gusev A,
- Trynka G,
- Reshef Y,
- Loh PR,
- et al
- Ni G,
- Moser G,
- Wray NR,
- Lee SH
- Bulik-Sullivan B,
- Finucane HK,
- Anttila V,
- Gusev A,
- Day FR,
- Loh PR,
- et al
- Gusev A,
- Lee SH,
- Trynka G,
- Finucane H,
- Vilhjálmsson BJ,
- Xu H,
- et al
- Randall JC,
- Winkler TW,
- Kutalik Z,
- Berndt SI,
- Jackson AU,
- Monda KL,
- et al
- Tang Y,
- Jin B,
- Zhou L,
- Lu W
- Boyle AP,
- Hong EL,
- Hariharan M,
- Cheng Y,
- Schaub MA,
- Kasowski M,
- et al
- Nikpay M,
- Šeda O,
- Tremblay J,
- Petrovich M,
- Gaudet D,
- Kotchen TA,
- et al
- Boyle EA,
- Li YI,
- Pritchard JK