Calculation of the Framingham Risk Score was based on the methods described by D'Agostino et al. (2008). Briefly, D'Agostino et al. (2008) developed a risk score from participants in the Framingham Heart Study to assess CVD (defined as a coronary death, myocardial infarction, coronary insufficiency, angina, ischemic stroke, hemorrhagic stroke, transient ischemic attack, peripheral artery disease, or heart failure) risk within the next 10 years. In the present study, this same method was used to assign risk scores to the current set of participants to be used as the phenotypic endpoint. These scores were calculated for men and women [separately] using the predictors of: (1) age, (2) HDL, (3) total cholesterol, (4) systolic blood pressure, (5) treatment status of systolic blood pressure, (6) smoking status, and (7) diabetic status. See D'Agostino et al. (2008) for a detailed description of the risk score calculation method 20. The calculated regression coefficients for risk score are presented in table 2.

Complete details on the available data, including assessment protocols can be found in the Framingham SHARe documentation. We note here that blood pressure was measured twice at each examination and the mean of these two measures was used when the participant had both scores. Cholesterol levels were determined by standard enzymatic methods while smoking and hypertensive treatment status was evaluated by participant and physician report. Diabetic status was defined as having definite diabetes (either being treated for diabetes or having a blood sugar reading of greater than or equal to 200 mg/dL).

For the original cohort, age, total cholesterol, blood pressure, smoking and diabetic status were assessed at Exam 1. HDL levels were not available for this cohort until Exam 9, making this measure an unusable predictor for risk score of this cohort. Therefore, a mean HDL level of all remaining participants from Exam 9 was calculated and applied to the original cohort individuals in conjunction with their Exam 1 scores. All predictors in the offspring and 3^rd generation cohorts were calculated from their first exam. With the exception of missing HDL levels in the original cohort, risk scores were only calculated for individuals with complete predictor information from the first exam.

Genome-wide genotypes and detailed clinical data on subjects have been made available to researchers through the SHARe (SNP-Health Association Resource) project. Unfiltered genotype data used in our study contained 9236 individuals genotyped for 500,568 SNPs (from the Affymetrix 500 K mapping array). We used PLINK, the whole genome association analysis toolset, in combination with R statistical computing software to perform quality control procedures 1021.

Subjects were excluded if genotyping rates were less than 95%. Individuals were also excluded if the predicted sex based on X-chromosome genotypes did not match the recorded sex. Evidence of non-random genotyping batch effects was inferred from identity-by-missingness (IBM) clustering of subjects. A batch covariate was generated for each subject and used to control for any batch effects. Subjects who were outliers with respect to estimated heterozygosity, those greater than 3 standard deviations from the mean, were excluded. All close relatives of individual subjects, based on mean identity-by-descent (IBD; PIHAT in PLINK) values indicating relatedness of less than 2^nd degree relatives, were excluded from the sample. All subjects with an identity-by-state (IBS) genetic distance from the sample mean of more than 3 standard deviations were considered outliers with respect to genetic ancestry and were pruned from the sample. This was also confirmed through visual inspection of Multidimensional scaling (MDS) plots. Markers were excluded if (1) genotyping rates were less than 95%, (2) minor allele frequencies were less than 0.01, and (3) if p-values from the Hardy-Weinberg Equilibrium (HWE) test were less than 1 × 10^-4. Evidence of non-random genotyping failure was examined, as inferred by flanking haplotypic background and significant markers were excluded (PLINK mishap test, p < 1 × 10^-6). We also removed individuals who had missing values for any covariates or phenotypic data. After quality controls, the remaining sample consisted of 1772 individuals genotyped on 258891 SNPs.

To minimize the chance that our polygenic heritability findings could be due to confounders (e.g., SNP batch effects) we used additional, very stringent quality control procedures for this analysis, as suggested in 11. Only those quality control methods that differ from those detailed above are listed below; otherwise all previously mentioned quality control procedures and generation of covariates were the same.

Markers were excluded if p-values from HWE tests were less than .05. Evidence of non-random genotyping failure was examined, as inferred by flanking haplotypic background and significant markers were excluded (PLINK mishap test, p < 1 × 10^-10)11. Subjects were excluded if genotyping rates were less than 99%. All close relatives of individual subjects, based on pairwise IBD (PIHAT) values greater than .025, were excluded from the sample when the GCTA software package was used to estimate heritability. We also removed individuals who had missing values for any covariates or phenotypic data. After quality controls, the remaining sample consisted of 1270 individuals genotyped on 250378 autosomal SNPs.

The polygenic heritability method involves first calculating the pairwise relatedness from SNPs between all individuals in a sample. When using the GCTA software package, any subjects that were related to any other subject with a degree greater than .025 IBD were dropped from the analysis altogether. This resulted in a significant loss of sample size (from 6725 to 1270) and thus a large reduction in power and an increase in standard error of the heritability estimate. To circumvent this large drop in sample size, we utilized an additional method of estimating heritability from pairwise relatedness based on a Haseman-Elston (H-E) regression framework 21. In a traditional H-E regression analysis, genetic relatedness is based on knowledge of pedigree and is regressed on trait difference between twins (either MZ or DZ). This generates an estimate for how much genetic similarity predicts phenotypic similarity. Our method uses a conceptually similar approach, but instead of using pedigree information, SNPs are used to determine the degree of genetic similarity between unrelated individuals. Using H-E regression, only those individual pairwise comparisons > .025 IBD (rather than entire subjects) were removed. This allows for the genetic relatedness information from all comparisons between subjects (other than those between close relatives) to be utilized in the analysis. Due to the non-independence between pairwise comparisons (the pairwise relationships from the same person are not independent of one another), we used bootstrapping to determine all significance levels and confidence intervals for H-E regression results. After quality controls, the sample used in the H-E regression analysis consisted of 6725 individuals genotyped on 250378 autosomal SNPs.

Initially a traditional GWAS analysis was performed on all SNPs using Framingham risk score as the target phenotype. Using the PLINK software package (v1.07) with the linear models option, a multiple linear regression test was performed on all quality controlled SNP data using 1772 individuals 10. An additive mode of inheritance was assumed and empirical p-values were generated for association with the quantitative phenotype at each locus after controlling for age, sex, genotyping batch effects, and (to control for the effects of population stratification) the 20 most significant principal components. A Manhattan plot and a Quantile-Quantile (Q-Q) plot were used to visualize association results. Prior to the analysis, we adopted the genome-wide significance threshold of p < 5 × 10^-8 to account for multiple testing 23. SNP call intensities were examined to determine whether significant SNPs were aberrations caused by poor quality genotyping. Significant SNPs within single genes based on NCBI Build 36.3 were then examined in Entrez Gene and PubMed to determine if any of these genes had previous associations with phenotypes thought to be related to cardiovascular disease.

Our second approach used a predictive polygenic model that effectively estimates the ability of common variants to predict the phenotypic score. Polygenic inheritance is expected to apply to many complex traits, CVD not withstanding 24. Thousands of small effects that are indistinguishable from background noise when performing a traditional GWAS can collectively account for a large proportion of the risk variation. Given the available sample size, we used a 10 fold cross-validation technique to minimize any effects of sampling error on polygenic prediction estimates 25. Methods of cross-validation are commonly used when the intent is to estimate how accurately a prediction model will perform with minimal bias due to the sampled data 26. This method involved partitioning the data from subjects into 10 discovery and test sets respectively. Each discovery set consisted of 90% of the individuals (N = 1575) from the total data set, while each test set consisted of the remaining 10% of subjects. Each of the 10 test sets was composed of a different 10% of the sample, thus no overlap between any of the test sets existed. For each discovery set, a GWAS analysis was performed 27. Using the results from this discovery phase, subsets of SNPs composed of those with p-values in the ranges of > .9-1, > .8-.9, > .7-.8, > .6-.7, > .5-.6, > .4-.5, > .3-.4, > .2-.3, > .1-.2, > 0-.1 were generated 13. The rationale behind grouping subsets of SNPs in ascending 0-.1 p-value bins is to verify the assumptions of a model where many SNPs of small effect are major contributors to CVD risk. With the current sample size, power to detect variants of small effects is individually very low, and by relaxing significance thresholds, power to detect common variants with small effect sizes increases (with a concomitant rise in number of false associations). SNPs from the discovery set in the lower p-value bins should on average have more predictive ability than those in the higher p-value bins, due to the higher ratio of true associations to type I errors 27.

Using the corresponding independent test set (the remaining 10%, N = 175), a genotypic risk score profile was assigned to each individual that represents the net predictive effect of all reference alleles possessed by a given individual. These genotypic risk scores were calculated using PLINK's SNP scoring routine. Each individual's score was calculated by summing the number of reference alleles at each locus multiplied by the value of the Beta for that respective SNP, which were generated from the initial survey of the discovery set 101327. Linear regression was then used to assess the relationship between risk of cardiovascular disease and the constructed polygenic risk scores in the test group 26. This same procedure was iterated for each of the 10 sample hold outs, allowing an unbiased estimate (in terms of r²) of the ability of common SNPs to predict the Framingham Risk Score.

Our third approach used two methods (restricted maximum likelihood and H-E regression) to estimate the proportion of the variance in the Framingham Risk Score that is explained by the common SNPs from the FHS data set. Instead of trying to predict heart disease risk from genotype, we examined if similarity in phenotype was significantly predicted by similarity of genotype. We examined if those subjects whose genomes are more similar than is expected by chance also have phenotypes that are more similar than is expected by chance. Using the GCTA software package 28, a genetic relationship matrix was generated based on shared combinations of SNPs and weighted by allele frequency. Then a linear model was fit to these genetic relationships using either a) a restricted maximum likelihood (REML) estimation maximization algorithm, as implemented in GCTA, or b) a linear regression model based on a H-E regression framework to estimate the variance explained by common SNPs 28.