Integrating Neuroimaging, Multi-omics, and Clinical Data in Complex Disease

摘要 生物医学信息学的快速发展产生了大量的高维数据集（“大数据”）， 范围从临床信息和医学成像到基因组序列数据。规模和复杂程度 这些数据集有很大的希望，但也存在重大挑战。为了充分利用 大数据的信息量，迫切需要找到有效的方法来整合来自 不同层次的信息技术。迄今为止的现有数据整合办法和方法 有几个重要的局限性。在这个项目中，我们提出了新的统计方法和策略， 集成神经成像、多组学和临床/行为数据集。增强联想力 分析相比，现有的方法，我们提出了一种新的多表型多变量关联 可以评估感兴趣的基因或区域中常见和罕见变异的累积效应的方法， 结合先前关于多表型结构的生物学知识， 在多个表型之间，并且对于高维表型是计算高效的。提高 临床结果的预测，我们提出了一种新的机器学习策略，可以整合 将多模态神经成像和多组学数据合并到数学模型中，并且可以将先验 生物学知识，以确定与临床结果相关的基因组相互作用。正在进行的 阿尔茨海默病神经成像倡议（ADNI）和印第安纳州记忆和衰老研究（IMAS）项目 作为试验台，提供了一个独特的机会来评估/验证所提出的方法。具体目标：目标1： 开发强大的统计方法，用于多表型之间关联的多变量检验 和单个遗传变异或一组变异（常见和罕见），并开发 用于整合神经影像学、遗传学和临床数据的中介分析方法， 通过神经影像学表型介导的间接遗传效应对临床结局的影响;目的2： 一种新的多变量模型，使用机器学习将多组学和神经成像数据结合起来， 预测患病个体或患病高风险个体的策略，并制定 一种新的多变量模型，它结合了先前的生物学知识，以确定相关的基因组相互作用。 目的3：使用ADNI的真实的数据评价和验证所提出的方法 目标4：传播和支持方便用户的公开软件， 有效地实现了所提出的方法。与公共卫生的关系：阿尔茨海默病 (AD)作为一个例子是一种越来越常见的进行性神经退行性疾病， 疾病改善治疗。提出的多变量方法可能有助于确定新的诊断方法， AD的生物标志物和治疗靶点。识别AD的新易感基因座/生物标志物， 重要的影响，获得更深入的分子机制，潜在的AD。

ABSTRACT Rapid progress in biomedical informatics has generated massive high-dimensional data sets (“big data”), ranging from clinical information and medical imaging to genomic sequence data. The scale and complexity of these data sets hold great promise, yet present substantial challenges. To fully exploit the potential informativeness of big data, there is an urgent need to find effective ways to integrate diverse data from different levels of informatics technologies. Existing approaches and methods for data integration to date have several important limitations. In this project, we propose novel statistical methods and strategies to integrate neuroimaging, multi-omics, and clinical/behavioral data sets. To increase power for association analysis compared to existing methods, we propose a novel multi-phenotype multi-variant association method that can evaluate the cumulative effect of common and rare variants in genes or regions of interest, incorporate prior biological knowledge on the multiple phenotype structure, identify associated phenotypes among multiple phenotypes, and be computationally efficient for high-dimensional phenotypes. To improve the prediction of clinical outcomes, we propose a novel machine learning strategy that can integrate multimodal neuroimaging and multi-omics data into a mathematical model and can incorporate prior biological knowledge to identify genomic interactions associated with clinical outcomes. The ongoing Alzheimer's Disease Neuroimaging Initiative (ADNI) and Indiana Memory and Aging Study (IMAS) projects as a test bed provide a unique opportunity to evaluate/validate the proposed methods. Specific Aims: Aim 1: to develop powerful statistical methods for multivariate tests of associations between multiple phenotypes and a single genetic variant or set of variants (common and rare) in regions of interest, and to develop methods for mediation analysis to integrate neuroimaging, genetic, and clinical data to test for direct and indirect genetic effects mediated through neuroimaging phenotypes on clinical outcomes; Aim 2: to develop a novel multivariate model that combines multi-omics and neuroimaging data using a machine learning strategy to predict individuals with disease or those at high-risk for developing disease, and to develop a novel multivariate model incorporating prior biological knowledge to identify genomic interactions associated with clinical outcomes; Aim 3: to evaluate and validate the proposed methods using real data from the ADNI and IMAS cohorts; and Aim 4: to disseminate and support publicly available user-friendly software that efficiently implements the proposed methods. RELEVANCE TO PUBLIC HEALTH: Alzheimer's disease (AD) as an exemplar is an increasingly common progressive neurodegenerative condition with no validated disease modifying treatment. The proposed multivariate methods are likely to help identify novel diagnostic biomarkers and therapeutic targets for AD. Identifying new susceptibility loci/biomarkers for AD has important implications for gaining greater insight into the molecular mechanisms underlying AD.