Abstract:
Blood lipid subfractions (total cholesterol, HDL cholesterol, LDL cholesterol, and triglycerides) are highly heritable, biomarkers and therapeutic targets of coronary artery disease, the leading cause of death. While the heritability for any lipid trait is 40-60%, the proportion we can explain by specific variants is very small. Our novel framework uses a multivariate adaptive shrinkage (mashR) model that allows the e!ect of a SNP to be modeled as a mixture of multivariate normal distributions across lipid subfractions, thus taking advantage of any boost in power gained by sharing among conditions. We use an empirical Bayes approach to learn the relative frequency of each pattern from the data to appropriately nudge the posterior estimates of the e!ect in accordance with the overall patterns observed in the data. In addition to reproducing all 318 loci recently significantly associated with blood lipid subfractions, in 297,626 ethnically diverse participants of the Million Veterans Program (MVP) with mashR, we identified 4,689 novel loci associated with lipids with summary-level data using mashR. We increase the total loci linked to lipids by 15-fold. In joint replication, 2129 of 5007 total loci (43%) detected in 297,626 MVP participants replicated in 330,000 UK Biobank participants, and 2129 of 2174 total loci (98%) detected in UK Biobank replicated in MVP. We show that associated variants are enriched in high-ranking eQTLs from the GTEx analyses in adipose and liver tissues. Finally, we provide an exploration of co-localization to better characterize the molecular function of identified targets. Lastly, we leverage our estimates across loci by assembling our posteriors as covariates in polygenic risk score construction. In summary, using a Bayesian approach, we significantly improve power for discovery using summary-level data.Our observations have broad implications for gene discovery and e!ect estimation, including for applications such as polygenic risk scores.