Transethnic Transferability of a Genome-wide Polygenic Score for Coronary Artery Disease

Genome-wide polygenic scores (GPS) integrate information from millions of sites of common DNA variation into a single metric of inherited susceptibility. When applied to coronary artery disease (CAD), we and others have noted three key observations: (i) GPS_CAD identifies a subgroup of the population at substantially increased risk not reliably identified using traditional risk factors; (ii) GPS_CAD can be quantified from birth, yet demonstrates similar or better discrimination power as clinical risk factors ascertained in middle age, and (iii) individuals with high GPS_CAD derive the greatest risk reduction from lipid-lowering therapy.¹

Despite growing evidence for integration of polygenic scores into routine clinical practice, the Eurocentric bias in genome-wide association studies – needed as input to polygenic score derivation – results in reduced performance of most polygenic scores in non-European ancestries.

To further quantify the transethnic portability of GPS_CAD, we assembled a meta-analysis of 30,365 CAD cases and 482,063 controls across five ancestral groups derived from six studies. First, the UK Biobank, a cohort study of middle-aged individuals that included 22,250 cases and 441,490 controls. Second, two hospital-based biobanks in the United States, the Partners Biobank with 2,772 cases and 15,124 controls, and the BioMe biobank with 2,824 cases and 21,547 controls.² Third, two ancestry-specific case-control studies, the Bangladesh Risk of Acute Vascular Events (BRAVE) study of early-onset myocardial infarction with 247 cases and 244 controls, and the TaiChi consortium in Taiwan with 288 cases and 457 controls.^3,4 Fourth, 1,984 cases with early-onset myocardial infarction from the hospital-based VIRGO study and 3,201 controls free of CAD from the MESA prospective cohort study.⁵ CAD was defined as myocardial infarction and/or history of coronary revascularization in all studies, with some variability across studies based on study design or available data as detailed previously.^1–5

Participants within each study provided informed consent to study investigators, with the meta-analysis approved by the Mass General Brigham institutional review board. UK Biobank data is made available to qualified researchers via application (http://www.ukbiobank.ac.uk/register-apply). Data for the VIRGO, MESA, TaiChi, and BRAVE studies have been uploaded to the database of Genotypes and Phenotypes (dbGaP) under accession numbers phs001259, phs001416, phs001487, and phs001398. Data from BioMe and Partners Biobank may be available via outreach to study investigators.

To quantify the relationship of GPS_CAD with disease, we first computed the GPS_CAD in each study participant based on either genotyping array and imputation (UK Biobank, Partners Biobank, and BioMe) or whole-genome sequencing (TaiChi, BRAVE, VIRGO/MESA). Sample and variant quality control was performed separately in each study as previously described.^1,2,5 Ancestral subgroups were based on concordance between self-report and genetically-defined ancestry, excluding individuals with discordant self-report and genetically-determined ancestry.^2,5 GPS_CAD was calculated using the raw weights for ~6.6 million variants (available at: www.broadcvdi.org).¹ We calculated the association of GPS_CAD with CAD using a logistic regression model adjusted for the first four principal components of ancestry. We then performed a meta-analysis of the odds ratio per standard deviation increment of GPS_CAD within ancestral groupings for each study (Figure).

Figure.

Association of a genome-wide polygenic score with coronary artery disease in five ancestry groups. In each cohort, the polygenic score is residualized to the first four principal components of ancestry (to adjust for the effect of ancestry on the score). In each ancestry of each study, the odds ratio per standard deviation is calculated using a logistic regression model adjusted for the first four principal components of ancestry (to adjust for the effect of ancestry on disease). Sensitivity analyses adjusting for sex and the first four principal components of ancestry, or age, sex, and the first four principal components of ancestry demonstrated similar results. A random-effects meta-analysis of odds ratio per standard deviation increment was performed. The p-value for subgroup differences was 0.005 (Cochran-Mantel-Haenszel test). In pairwise ancestry group comparisons, the effect estimate in African ancestry group was significantly lower than each of the other ancestry groups, European (p=0.002), South Asian (p=0.02), East Asian (0.001), and Hispanic (0.003). None of the other ancestry group pairwise comparisons were statistically significant (p>0.05). The random effects model results for each ancestry group showed an I² of 95.4%, 63.7%, 0%, 0%, and 0% for European, African, South Asian, East Asian, and Hispanic, respectively. Abbreviations: AJ Ashkenazi Jewish, CI confidence interval, SD standard deviation.

GPS_CAD was robustly associated with CAD within each ancestral group, but significant variability in effect size was noted (p-value=0.005). Consistent with prior reports, we observe an odds ratio per standard deviation (OR_SD) of 1.60 (1.44–1.78) in individuals of European ancestry. Similar effect sizes were noted in East Asian (OR_SD = 1.66 [1.47–1.86]), South Asian (OR_SD =1.47 [1.36–1.59]), and Hispanic (OR_SD =1.52 [1.43–1.62]) ancestries. Among individuals of African ancestry, the GPS_CAD remained associated, but effect size was significantly reduced (OR_SD =1.25 [1.12–1.40]).

The transethnic variability – most notable for a decreased effect size in individuals of African ancestry – is likely attributed to four factors: (i) use of non- causal variants that – due to variable linkage disequilibrium patterns – may not correlate with the causal signal across ancestries; (ii) allele frequency differences or population-specific risk variants that have not been optimally modeled; (iii) varying genetic architectures, as can occur with gene-gene or gene-environment interactions; (iv) different heritabilities, as can occur with population-specific environmental exposures. More diverse genome-wide association studies remain a top priority for the scientific community. In parallel, ongoing efforts seek to improve performance of polygenic scores in individuals of African ancestry with existing data, such as improving fine mapping to identify causal variants, optimizing methods that combine data from different ancestries, and including other sources of information (risk factors or other ‘omics data).

This study should be interpreted in the context of potential limitations. Within each ancestral group, studies were somewhat variable in their design and does not consider non-genetic risk factors. Additionally, we quantify relative risk associated with GPS_CAD using odds ratios, but epidemiological differences in CAD risk across ancestral groups influence estimates of absolute risk. For example, despite a lower relative risk related to a polygenic score, a greater number of individuals of Africa ancestry may benefit from reclassification in the context of higher absolute event rates.

Nonstandard Abbreviations and Acronyms

GPS_CAD

Genome-wide polygenic score for coronary artery disease

CAD

Coronary artery disease

OR_SD

Odds ratio per standard deviation increment

BRAVE

Bangladesh Risk of Acute Vascular Events study

MESA

Multi-Ethnic Study of Atherosclerosis

VIRGO

Variation In Recovery: role of Gender on Outcomes of young acute myocardial infarction patients study

BioMe

BioMe biobank of Icahn School of Medicine at Mount Sinai

SNP

Single nucleotide polymorphism

AJ

Ashkenazi Jewish