Collaborative Research: Statistical Methods for Integrated Analysis of High-Throughput Biomedical Data

技术进步导致了医学中高通量“组学”数据的迅速扩散，这些数据掌握着临床上有效的个性化医学的关键。为了实现这一目标，迫切需要统计和计算工具来挖掘这些数据并发现生物标记物、药物靶标、被破坏的疾病网络和疾病亚型。然而，有两个主要因素使这类统计工具的开发具有挑战性。首先，许多高通量基因组技术产生各种不同的异质数据，包括连续数据(微阵列、甲基化阵列)、计数数据(RNA测序)和二进制/分类数据(SNPs、CNV)。这些不同的数据集并不总是满足标准高维统计模型强加的典型分布假设。其次，为了让科学家利用他们的所有数据并了解疾病的完整分子基础，这些不同的组学数据集需要合并到一个单一的多变量统计模型中。这项提议试图用一个新的统计框架来解决这两个问题，以便对同一组受试者测量的多组高维数据进行综合分析。关键统计方法使用指数族分布理论来概括主成分分析(PCA)和图形模型这两个基本的高维统计框架，从而联合分析转录、表观基因组和功能基因组数据。这项研究将应用于高通量的癌症基因组数据，并导致新的方法：(A)发现分子癌症亚型及其基因组特征；(B)建立疾病的整体网络模型。通过利用所有可用不同类型的基因组生物标记物的信息，拟议的方法将有可能做出对个性化医学至关重要的科学发现。这项拟议的工作还将广泛适用于整合多组“组学”数据，包括基因组学、蛋白质组学、代谢组学和成像。除了医学，理论框架和统计方法将在指数族理论、统计学习和新兴的综合分析领域取得重大进展，并在工程和安全等其他学科中具有广泛的适用性。所有成果将通过出版物、会议和开源软件传播；这项研究还将为博士和博士后学者提供培训和教育机会。

Technological advances have led to a rapid proliferation of high-throughput "omics" data in medicine that hold the key to clinically effective personalized medicine. To realize this goal, statistical and computational tools to mine this data and discover biomarkers, drug targets, disrupted disease networks, and disease sub-types are urgently needed. There are, however, two primary factors which make the development of such statistical tools challenging. First, many high-throughput genomic technologies produce varied heterogeneous data, which include continuous data (microarrays, methylation arrays), count data (RNA-sequencing), and binary/categorical data (SNPs, CNV). These varied data sets do not always satisfy typical distributional assumptions imposed by standard high-dimensional statistical models. Second, in order for scientists to leverage all of their data and understand the complete molecular basis of disease, these varied omics data sets need to be combined into a single multivariate statistical model. This proposal seeks to address these two issues with a new statistical framework for integrated analysis of multiple sets of high-dimensional data measured on the same group of subjects. The key statistical approach uses the theory of exponential family distributions to generalize two foundational high-dimensional statistical frameworks, principal components analysis (PCA) and graphical models, so as to jointly analyze transcriptional, epi-genomics and functional genomics data. This research will be applied to high-throughput cancer genomics data and lead to new methods to (a) discover molecular cancer sub-types along with their genomic signatures and (b) build a holistic network model of disease. By leveraging information across all the different types available of genomic biomarkers, the proposed methods will have the potential to make scientific discoveries critical for personalized medicine. The proposed work will also be broadly applicable to integrating multiple sets of "omics" data, including genomics, proteomics, metabolomics, and imaging. Beyond medicine, the theoretical framework and statistical methods will make significant advances in the theory of exponential families, statistical learning, and the emerging field of integrative analysis as well as have broad applicability in other disciplines such as engineering and security. All results will be disseminated through publications, conferences, and open-source software; this research will also provide training and educational opportunities for doctoral and postdoctoral scholars.