系统流行病学中多尺度高维因果中介分析的理论与统计方法研究

Systems epidemiology has highlighted the significance that we can open the “black box” from the exposure to the disease, by absorbing the multi-scale omics data generated from advanced high-throughput-omics technologies. Statistically, it is essential to decompose the effect of the exposure on the outcome and conduct the mediation analysis. However, current mediation analysis methods is not suitable and deviated much from the design of systems epidemiology. Given the omics data in“black box” is often multi-scale and high-dimensional, this project provide the theoretical framework of causal mediation analysis with multi-scale and high dimensional mediators. Using prior distribution to appropriately describe the different biological attributes behind of muiti-scale omics, the Bayesian variable selection has been incorporated into the high-dimensional mediation model, meanwhile, the posterior inclusion probability has been adopted to measure the uncertainty that one single variable is the true mediator. Then we can conduct the regular statistical inference and depicted the causal mediation network from the exposure to outcome. Various simulations and sensitive analysis are conducted to assess the model performance and robustness respectively. Furthermore, one real data set related with metabolic traits has been analyzed to evaluate the practicability. It is expected that this project can construct statistical theory and methods for mediation analysis with multi-scale and high dimensional mediators, accompanying with user-friendly software package. It will be helpful to derive disease pathway, design drug targets as well as develop and evaluate intervention strategies.

系统流行病学在设计上突出强调吸收现代高通量组学技术，在群体水平上借助多尺度组学标记，以期打开暴露到疾病的“黑盒子”。将暴露对疾病的效应细化分解，进行因果中介分析是其统计方法核心，然而，当前中介分析策略均不能满足其设计需求。本项目着眼于“黑盒子”承载的组学标记具有“多尺度”和“高维”双重特征，提出了多尺度高维中介变量情形下因果中介分析理论框架，用不同的先验分布表征和刻画多尺度组学标记所蕴含的不同生物学属性，将贝叶斯变量选择嵌入高维中介模型，用后验包含概率量化组学标记为真实因果中介的不确定性，同时实现正则化统计推断，进而勾画出从暴露到疾病的因果中介网络。统计模拟评价模型的科学性和优劣性，敏感性分析评价其稳健性，实例分析代谢疾病系统流行病学数据验证其实用性，以期构建“系统流行病学多尺度高维因果中介分析的理论与统计方法体系及其软件包”，为推断致病通路、设计药物靶点、制定及评估干预策略提供新方法。

系统流行病学在设计上突出强调吸收现代高通量组学技术，在群体水平上借助多尺度组学标记，以期打开暴露到疾病的“黑盒子”，对多组学数据整合并研究因果关联推断方法是其统计方法核心。本项目初步探索了相应的统计方法体系，并开展了一系列数据分析，主要完成4项工作：1）在因果孟德尔随机化网络框架下，基于联合似然理论，构建跨组学数据整合的PMR-Egger模型，不仅能控制SNP间LD关系而且能矫正水平多效性，实现了在水平多效性存在时，检验基因表达对性状的因果效应。进一步，考虑到很多复杂性状往往具有共同的遗传基础，且对多性状联合分析可以充分利用性状之间的相关性，可提高检验效能，将PMR-Egger方法推广，提出了多性状新型moPMR-Egger模型；2）着眼于候选工具变量所承载的难以界定的多重效应角色（垂直多效性、相关水平多效性、不相关水平多效性等），分层次、分脉络对其分别加以刻画，将变量选择嵌入模型，用后验概率量化其不确定性，实现多重效应的自适应定位，校正水平多效性，基于贝叶斯渐近统计理论，沿抽样逼近似然函数思路，突破高维积分求解瓶颈，实现联合似然一体化推断，提出了工具变量自适应定位的两样本MR联合似然新型统计方法；3）构建了跨组学无向网络（及特定通路）的回归模型, 引入点互信息测量两个网络节点之间的连接强度，构建了基于互信息的组学无向网络回归模型，将其进一步推广到跨组学TWAS整合框架，依托两阶段推断步骤，先后构建了分别针对连续结局、二分类和有序多分类结局的TWAS网络回归模型；4）依托上述方法学理念，系统探索了路易体痴呆、阿尔茨海默症和帕金森病的共同遗传基础、完成了幼年特发性关节炎和特发性肺纤维化（IPF）的跨组学数据整合分析，并探索了出生体重和成年患慢性肾病、甲状腺功能减退症（甲减）和IPF、茶叶摄入和二型糖尿病、甲状腺功能和血脂水平、甲减和原发性胆汁性肝硬化间的因果关联关系。