Robust Methods for Polygenic Analysis to Inform Disease Etiology and Enhance Risk Prediction

摘要 现代全基因组关联研究已经明确表明，复杂的性状是极其复杂的。 多基因的，每个个体特征可能涉及数千到数万个遗传变异。在 在这个项目中，我们将开发一系列新的方法来大规模利用多基因信号的力量。 GWAS告知疾病病因并改进风险预测模型。在（目标1）中，我们将开发方法 用于在GWAS中进行与各种群体遗传相关的关联信号的富集分析， 和基因组的功能基因组特征。我们建议对效应量分布进行建模 使用灵活的正态混合模型与标记的全基因组面板相关联，其中类 标记的成员是根据各种基因组“协变量”概率建模的。推断 模型和基本参数将进一步利用在一个统计贝叶斯框架，以得出多基因 用于遗传风险预测的风险评分（PRS）。在（目标2）中，我们将开发新的方法， 随机化分析，一种工具变量分析形式，用于调查因果关系 风险因素和健康结果之间的关系。我们将利用灵活的模型来分析双变量效应量分布 在成对的性状之间，允许遗传相关性从因果关系和非因果关系中产生。 我们提出了一个解决方案的复杂问题的因果效应估计的建议框架下 在某些类型的残差分布中使用创新的“尖峰检测”方法。在（目标3）中，我们 将开发新的方法，以提高使用PRS的基因-环境相互作用分析的能力， 对照研究。我们将开发可以利用各种自然假设的回顾性方法 关于PRS的分布，包括正态性及其与环境暴露的独立性， 可能取决于其他因素，在潜在人群中。我们将把建议的方法应用于 对现有的GWAS数据集进行大规模分析，以获得各种各样的性状，并期望在这些性状中产生新的 关于遗传易感性机制的科学观察，流行病学的因果关系 关联，基因-环境相互作用的性质和遗传风险预测的实用性。

Abstract Modern genome-wide association studies have unequivocally demonstrated that complex traits are extremely polygenic, with each individual trait potentially involving thousands to tens of thousands of genetic variants. In this project, we will develop a series of novel methods to harness the power of polygenic signals in large GWAS to inform disease etiology and improve models for risk prediction. In (Aim 1), we will develop methods for conducting enrichment analysis of association signals in GWAS in relationship to various population genetic and functional genomic characteristics of the genome. We propose to model effect-size distributions associated with whole genome panel of markers using flexible normal-mixture models, where class memberships of the markers are modelled probabilistically in terms of various genomic “covariates”. Inferred models and underlying parameters will be further utilized in an empirical-Bayes framework to derive polygenic risk-scores (PRS) for genetic risk prediction. In (Aim 2), we will develop novel methods for Mendelian randomization analysis, a form of instrumental variable analysis, for the investigation of causal relationships between risk-factors and health outcomes. We will utilize flexible models for bivariate effect-size distributions across pairs of traits, allowing for genetic correlation to arise from both causal and non-causal relationships. We propose a solution to the complex problem of estimation of causal effects under the proposed framework using an innovative method for “spike detection” in the distribution of certain types of residuals. In (Aim 3), we will develop novel methods to enhance the power of gene-environment interaction analysis using PRS in case- control studies. We will develop retrospective methods that can take advantage of various natural assumptions about the distribution of PRS, including normality and its independence from environmental exposures, possibly conditional on other factors, in the underlying population. We will apply the proposed methods to conduct large scale analysis of existing GWAS datasets for a wide variety of traits and expect to make novel scientific observations regarding mechanisms of genetic susceptibility, causal basis for epidemiologic associations, nature of gene-environment interactions and utility of genetic risk prediction.