Statistical Methods for Integrative Analysis of Genomics and Proteomics Data

描述（由申请人提供）：肿瘤是复杂的生物系统。没有单一类型的分子方法完全阐明肿瘤行为，需要在包括基因组学和蛋白质组学在内的多个水平上进行分析。因此，现在在全基因组范围内收集来自许多来源的不同类型的数据，包括：DNA拷贝数改变，mRNA表达，蛋白质表达测量和许多其他数据。然而，如果没有有效的统计和计算方法，这些研究中的生物医学信息的全部范围就无法实现。因此，这项研究的长期目标是开发创新方法，联合建模这些不同类型的数据，以帮助揭示基因和蛋白质相互作用的大规模组织。为了应对这一挑战，该提案从目标1开始，通过开发新的统计和计算方法来识别DNA/RNA/蛋白质相互作用。我们建议使用图形模型开发的工具，研究基因/蛋白质与各种条件相关性之间的条件依赖关系。目的2提出新的方法来整合相互作用网络与疾病表型，以改善生物标志物识别和临床结果预测。我们将推导出与疾病发生/进展相关的基因/蛋白质模块，并使用增强程序将模块信息纳入预测模型。稀疏回归技术和适当的平滑正则化将用于处理高维数据，并解释两个目标的局部相关性。该提案使用两项乳腺癌研究作为激励性例子。但这里开发的工具可以很好地推广到其他疾病。 这项研究的成功将大大改善大规模整合研究的统计方法，从而有助于提高对基因组/蛋白质组学改变对肿瘤生长和进展的贡献的机制理解，并促进更有效的分子诊断和预后测试的发展。来自两项乳腺癌研究的数据将与广泛的模拟实验一起使用，以测试和改进现实世界应用的方法。

DESCRIPTION (provided by applicant): Tumors are complex biological systems. No single type of molecular approach fully elucidates tumor behavior, necessitating analysis at multiple levels encompassing genomics and proteomics. Therefore different types of data from numerous sources are now collected at a genome-wide scale, including: DNA copy number alterations, mRNA expression, protein expression measurements and many others. However, the full extent of biomedical information in these studies cannot be realized without effective statistical and computational methods. Thus, the long-term goal of this research is to develop innovative methods jointly modeling these different types of data to help uncover the large-scale organization of genes and proteins interacting. To tackle this challenge, this proposal begins in Aim 1 by developing new statistical and computational methods for identifying DNA/RNA/Protein interactions. We propose to use tools developed for graphics models and study conditional dependencies among genes/proteins with various conditional correlations. Aim 2 proposes novel approaches to integrate the interaction network with disease phenotypes to improve biomarker identification and clinical outcome prediction. We will derive modules of genes/proteins which are associated with disease initiation/progression, and use boosting procedures to incorporate the module information into the predictive models. Sparse regression techniques together with proper smooth regularization will be used to handle the high-dimensionality and to account for the local correlation in both aims. The proposal uses two breast cancer studies as motivating examples. But the tools develop here can be well generalized to other disease. Success of this research will result in substantially improved statistical methods for large-scale integration studies, and thus help to increase mechanistic understanding of the contribution of genomic/proteomics alterations to tumor growth and progression, as well as facilitate the development of more effective molecular diagnostic and prognostic tests. Data from the two breast cancer studies will be used together with extensive simulation experiments to test and refine the methodology for real-world application.