Estimation and inference in directed acyclic graphical models for biological networks

总结 随着生物技术的进步，生物医学研究由于高通量和高效率而变得更加复杂。 以基因组规模收集的高维数据。最重要的是， 基因上的遗传变异和基因对基因的调控关系。在这一点上，生物医学研究人员 可以识别因果单核苷酸多态性（SNP）和复杂性状和神经退行性疾病的基因 阿尔茨海默病（AD）等疾病的研究，以制定治疗策略。鉴于迫切需要了解- 面对这些疾病，特别是AD的进展和病因，PI建议制定统计和 用于精确估计和推断基因调控网络的计算工具，重点是AD和 其他复杂的特征。 该项目包括两个主要组成部分：基因调控网络的估计和推断， SNP作为工具变量（IV）。重点将放在因果网络重构和推理上 在可能存在无效IV和隐藏混杂因素的情况下， 在高维数据中，变量的数量可能超过样本大小。关于因果关系 网络重建，该项目将开发重建基因调控网络的新方法， 有向无环图描述了SNP（干预），基因和性状之间的因果关系， 如AD。该项目将开发基于改进的艾德似然比检验和 数据扰动方案来解释发现过程中涉及的不确定性。此外，它将重点关注 关于（1）多重（线性/非线性）因果关系的方向性和强度的假设检验，以及（2） 因果关系路径的存在。在计算方面，该项目将开发创新的方法， 解决大规模问题的算法在应用方面，基于重构的基因调控网络， 将首先确定AD的致病基因和AD的风险因素，如脂质，然后推断哪些风险因素是 （puzzle）与AD有因果关系。

Summary As biotechnology advances, biomedical investigations have become more complex due to high-throughput and high-dimensional data collected at a genomic scale. Of paramount importance is unraveling the regulatory roles of genetic variants on genes and gene-to-gene regulatory relationships. On this ground, biomedical researchers can identify causal Single-Nucleotide Polymorphisms (SNPs) and genes for complex traits and neurodegenerative diseases such as Alzheimer's disease (AD) to develop treatment strategies. Given the urgent need to under- stand the progression and etiology of these diseases, particularly AD, the PIs propose to develop statistical and computational tools for accurate estimation and inference of gene regulatory networks, with a focus on AD and other complex traits. The project consists of two major components: estimation and inference of gene regulatory networks with SNPs as instrumental variables (IVs). The main thrust will be on causal network reconstruction and inference with IVs as interventions in the possible presence of invalid IVs and hidden confounders, with particular effort on high-dimensional data, in which the number of variables may exceed the sample size. Concerning causal network reconstruction, the project will develop novel methods of reconstructing gene regulatory networks as directed acyclic graphs describing casual relationships among the SNPs (interventions), genes, and traits such as AD. The project will develop high-dimensional inferential tools based on modified likelihood ratio tests and a data perturbation scheme to account for the uncertainty involved in a discovery process. Moreover, it will focus on hypothesis testing on (1) the directionality and strength of multiple (linear/nonlinear) causal relations and (2) the presence of a pathway of causal relations. Computationally, the project will develop innovative methods and algorithms for large-scale problems. For application, based on the reconstructed gene regulatory networks, we will first identify causal genes for AD and AD's risk factors, such as lipids, then infer which of the risk factors are (putatively) causal to AD.