Broad clinical manifestations of polygenic risk for coronary artery disease in the Women's Health Initiative.

Background: The genetic basis for coronary artery disease (CAD) risk is highly complex. Genome-wide polygenic risk scores (PRS) can help to quantify that risk, but the broader impacts of polygenic risk for CAD are not well characterized.
Methods: We measured polygenic risk for CAD using the meta genomic risk score, a previously validated genome-wide PRS, in a subset of genotyped participants from the Women's Health Initiative and applied a phenome-wide association study framework to assess associations between the PRS and a broad range of blood biomarkers, clinical measurements, and health outcomes.
Results: Polygenic risk for CAD is associated with a variety of biomarkers, clinical measurements, behaviors, and diagnoses related to traditional risk factors, as well as risk-enhancing factors. Analysis of adjudicated outcomes shows a graded association between atherosclerosis related outcomes, with the highest odds ratios being observed for the most severe manifestations of CAD. We find associations between increased polygenic risk for CAD and decreased risk for incident breast and lung cancer, with replication of the breast cancer finding in an external cohort. Genetic correlation and two-sample Mendelian randomization suggest that breast cancer association is likely due to horizontal pleiotropy, while the association with lung cancer may be causal.
Conclusion: Polygenic risk for CAD has broad clinical manifestations, reflected in biomarkers, clinical measurements, behaviors, and diagnoses. Some of these associations may represent direct pathways between genetic risk and CAD while others may reflect pleiotropic effects independent of CAD risk.

Introduction

Coronary artery disease (CAD) is a complex phenotype, and the genetic basis for CAD risk is similarly complex[1]. To date, >200 loci have been implicated in CAD risk through genome-wide association studies (GWAS)[2,3]. These loci interact through a diverse set of biological pathways, and many loci have no apparent relevance to traditional risk factors for CAD. Furthermore, genetic variants that associate with CAD also associate with other phenotypes, suggesting extensive underlying pleiotropy[2,3]. The complexity of genetic risk for CAD is further highlighted by recent advances in the construction of polygenic risk scores (PRS). Contemporary scores that incorporate variants across the whole genome, including variants outside of known CAD loci, outperform scores that are constructed only from variants at known CAD loci[4,5]. Studying such genome-wide PRS for CAD may allow for improved understanding of the genetic basis for CAD risk and new insights into the implications polygenic risk beyond CAD.

One approach to assessing the impact of polygenic risk for CAD has been to measure associations between a CAD PRS and biobank-derived phenotypes[6,7]. A primary advantage of this approach is the large number of participants in such biobanks. However, a limitation of this method is lack of precision for some outcomes, particularly those inferred from electronic health records. Further, biobank studies have typically combined prevalent and incident disease and may have limited follow-up after enrollment. Thus, a complementary approach to biobank analyses is to examine well-phenotyped longitudinal cohorts.

Here, we seek to identify traits and outcomes associated with polygenic risk for CAD by taking advantage of the high-quality data collected as part of the Women’s Health Initiative (WHI). We aggregate data collected over ~25 years as part of either clinical trials or the observational study within WHI. We thus measure the association between polygenic risk for CAD and blood biomarkers, clinical measurements, clinical risk scores/questionnaires, self-reported medical history, and incident adjudicated outcomes related to cardiovascular disease, cancer, and death.

Methods

Study cohort

The main study cohort was selected from WHI. The design and recruitment strategy for WHI has been previously described[8,9]. Briefly, postmenopausal women aged 50 to 79 years were enrolled at 40 sites across the United States from 1993 to 1998. Each participant was enrolled into either a clinical trial (n = 68,132) or an observational study (n = 93,676). Two successive extension studies continued follow-up of consenting participants from 2005 to 2010 and from 2010 to the present. A subset of participants who were primarily non-Hispanic white by self-report have been previously genotyped as part of 6 ancillary GWAS (Supplementary Table 1). Participants from those 6 GWAS were considered for inclusion in this study. Because currently validated genome-wide PRS were developed in European populations and do not transfer well to non-European populations, we did not include cohorts of primarily non-European genotyped participants in this study. Subjects with a known or likely history of atherosclerotic cardiovascular disease (ASCVD) at enrollment were excluded (Supplementary Table 2). We used the UK Biobank for replication of select results. The UK Biobank cohort consisted of unrelated post-menopausal women of European ancestry with no history of MI or stroke at enrollment.

Genotyping and imputation

Genotyping was performed with early versions of Affymetrix and Illumina gene chips for five of the GWAS cohorts contributing to this study. For these five studies, harmonization and imputation to the 1000 Genome reference panel was previously performed as part of the WHI GWAS Harmonization and Imputation Project (https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000746.v3.p3). Participants of the sixth study were genotyped with the Oncochip, and we imputed these data to the 1000 Genome reference panel using the Michigan Imputation Server[10].

Main exposure

We used metaGRS, a previously developed genome-wide PRS for CAD, to estimate each participant’s genetic risk[5]. This score consists of ~1.7 million autosomal variants. Participants in our study cohort did not contribute to the GWAS used to construct this score. Each participant’s total score was calculated using Plink 2.0, and raw scores were then scaled to mean 0 and standard deviation 1. This standardize score was used as the primary exposure.

Phenotypes

Quantitative measurements were largely collected at enrollment and included laboratory values, clinical measurements, and clinical scores. For the small number of lab measurements not collected at baseline, we used the earliest available measurement. Lab outliers were removed by excluding the top 1% of values for each biomarker. For clinical measurements such as blood pressure, the mean value was used if serial measurements were available within one research clinic visit. Self-reported medical history, medication usage, social/behavioral history, and family history was obtained through questionnaires collected primarily at enrollment but also during annual follow-up mailings. Adjudicated outcomes assessed in this study include incident cardiovascular diseases, incident cancers, and death. Annual questionnaires were completed by participants or their proxies in order to identify hospitalizations, and for each hospitalization, medical records were obtained and adjudicated by physicians using standardized criteria[11]. Deaths were further ascertained through the National Death Index. For UK Biobank analyses, cancer diagnoses were extracted from the UK cancer registry. For each cancer, only first diagnoses after enrollment were considered as incident cases, and subjects with prevalent disease at enrollment were excluded.

Statistical analysis

We selected the largest subset of subjects with similar inferred genetic ancestry using principal components analysis in order to limit confounding by population substructure. We used linear and logistic regression to estimate associations between each trait/outcome and the CAD PRS per standard deviation increase in the PRS. For each of the adjudicated outcome, we appropriately censored subjects at the end of the follow-up time period where formal adjudication ended for the outcome. For death outcomes, we used Cox analysis with time zero being the time of enrollment. For each cause of death that was examined, non-cases were censored at time of death from another cause or time of last follow-up if not deceased.

Each model was adjusted for age at enrollment (or age at time of measurement for lab values), study type (clinical trial versus observational study), and genotyping platform. Associations with lipid-related labs, diabetes-related labs, and for blood pressure were additionally adjusted for self-reported cholesterol medication use, diabetes medication use, and hypertension medication use respectively. All associations with lab values were also adjusted for the assay version if more than one assay was used. For the analysis of self-reported outcomes, we compared three associations. First, we performed logistic regression using the main study cohort, adjusting for age at enrollment, study type, and genotyping platform. Second, we added an additional binary covariate to adjust for presence or absence of CAD at the last follow-up. Third, we analyzed the subset of participants with no CAD at follow-up (n = 18,044), adjusting for age at enrollment, study type, and genotyping platform. CAD at follow-up was determined using both self-report and adjudicated outcomes. Only outcomes with at least 100 cases among the CAD-free cohort were considered, resulting in a total of 128 self-reported qualitative variables. The logistic regression analysis of adjudicated cardiovascular outcomes and the Cox analysis of death outcomes were adjusted for smoking status, self-reported diabetes at baseline, systolic blood pressure, low-density lipoprotein cholesterol (LDL-C), and high-density lipoprotein cholesterol (HDL-C). For cancer outcomes, we assessed the association with and without adjustment for risk factors. The risk factor adjusted model include adjustment for smoking status, alcohol consumption, physical activity (MET-hours per week), Alternative Healthy Eating Index score, and body-mass index (BMI).

Using a phenome-wide association study framework, we consider statistical significance in three ways. Nominal significance is defined as a p-value ≤ 0.05. Where indicated, we also identify associations that are significant by a false-discovery rate (FDR) q-value ≤ 0.05, using the Benjamini and Hochberg method. Lastly, Bonferroni significance is defined as 0.05 divided by the number of association tests performed for the given analysis. For the association analysis of quantitative traits, Bonferroni significance was p-value ≤ 9.2 × 10−5 (0.05/546). For the association analysis of incident cardiovascular diseases, Bonferroni significance was p-value ≤ 0.003 (0.05/17).

We used published summary statistics from GWAS of CAD[2], breast cancer[12], and lung cancer[13] to estimate genetic correlations. For CAD, only variants with INFO score >0.9 were included. For breast and lung cancer, INFO score was not available, and thus only HapMap3 variants were included, as these variants are generally well imputed. We used ldsc (version 1.01) to perform genetic correlation analyses[14].

We performed two-sample Mendelian randomization using the MRBase tool with default settings[15]. We created a genetic instrument variable for CAD using the same GWAS as was used in the genetic correlation analysis[2]. We selected genome-wide significant SNPs that were determined to be independent using a clumping distance of 10 megabases with a linkage disequilibrium R2 threshold of 0.001. These were then harmonized to the summary statistics of each outcome, excluding palindromic SNPs and using proxies for missing SNPs only if the LD R2 was ≥ 0.9. For lung cancer, the instrument variable consisted of 125 SNPs, of which 1 SNP was proxied. For breast cancer, the instrument variable consisted of 124 SNPs, of which none were proxied. MRBase was used to perform inverse variance weighted, weighted median, and MR Egger studies. As additional sensitivity analysis, and to test for horizonal pleiotropy, we used MR PRESSO[16].

WHI analyses were performed using SAS 9.4 (SAS Enterprise). UK Biobank analyses, meta-analysis, Mendelian randomization, and plots were done with R version 3.5.1 (R Foundation, Vienna, Austria). All odds ratios (OR) and hazard ratios (HR) are reported as per standard deviation increase in the PRS.

Ethics statement

The WHI project was reviewed and approved by the Fred Hutchinson Cancer Research Center (Fred Hutch) IRB in accordance with the U.S. Department of Health and Human Services regulations at 45 CFR 46 (approval number: IR# 3467-EXT). Participants provided written informed consent to participate. Additional consent to review medical records was obtained through signed written consent. Fred Hutch has an approved FWA on file with the Office for Human Research Protections (OHRP) under assurance number 0001920. WHI data were accessed through the sponsorship of T. Assimes (WHI co-investigator) and with an approved proposal (MSID 3914). The UK Biobank data was accessed under Application Number 13721. All participants gave informed consent for participation in UK Biobank. The Research Ethics Committee reference for UK Biobank is 16/NW/0274. This study of pre-existing de-identified data was deemed not human subjects research by the Stanford IRB, and thus no further consent was obtained.

Table 1

Two-sample Mendelian randomization analyses testing for a causal association between the exposure of coronary artery disease and cancer outcomes.

Outcome	Method	OR (95% CI)	P Value
	Inverse variance weighted	0.95 (0.92–0.99)	0.009
	Weighted median	0.96 (0.92–1.01)	0.1
Breast cancer	MR Egger	0.97 (0.90–1.04)	0.4
	MR PRESSO raw	0.96 (0.93–1.00)	0.04
	MR PRESSO outlier-corrected	0.98 (0.95–1.02)	0.4
	Inverse variance weighted	0.92 (0.84–1.00)	0.05
	Weighted median	0.84 (0.75–0.94)	0.002
Lung cancer	MR Egger	0.77 (0.65–0.91)	0.003
	MR PRESSO raw	0.91 (0.84–0.99)	0.04
	MR PRESSO outlier-corrected	0.92 (0.86–0.99)	0.04