Statistical Methods for Genetic Risk Predictions across Diverse Populations

点击翻译按钮获取中文摘要

Project Summary The ultimate goal of the parent grant is to develop rigorous, efficient, and robust integrative modeling approaches for risk prediction across populations by capitalizing on the vast amount of publicly available GWAS summary data, abundant functional annotations, and a growing number of studies with participants from underrepresented populations. There are five specific aims in the parent grant, with the first three aims developing three complementary approaches for cross-population risk predictions, including: (Aim 1) a Bayesian approach (ME-Pred), along the line of our published work to incorporate either functional annotation information or multiple trait information, that explicitly models joint effect sizes from multiple populations and functional annotations; (Aim 2) an empirical Bayes approach (GWEB) that considers a more general and flexible effect size distribution and statistical inference that does not need a validation cohort for tuning some model parameters; and (Aim 3) a fast and robust Bayesian nonparametric method (SDPR) that is highly adaptive to different genetic architectures and is computationally efficient. Aim 4 and Aim 5 of the parent grant focus on implementation and applications of the developed tools to a number of studies, including the Generations Project jointly initiated by the Yale School of Medicine and the Yale New Haven Health System. We have assembled a team of investigators with expertise in statistical genetics, medical genetics, and high- performance computing to develop, implement, evaluate, and disseminate the proposed methods. For this diversity supplement project, we will consider risk prediction in admixed individuals, a topic related to but covered by the parent grant. In addition, we will consider the applications of our methods to evaluate disease risk for participants in the All of Us Research Program with a substantial number of underrepresented individuals where methods tailored for admixed subjects may provide significantly improved predictions. This supplement will not only provide an excellent training opportunity, but will also develop new tools for disease risk predictions.