High-dimensional unsupervised learning, with applications to genomics

高维无监督学习及其在基因组学中的应用

• 批准号: 8708556
• 负责人: Daniela Witten
• 金额: $ 36.38万
• 依托单位: UNIVERSITY OF WASHINGTON
• 依托单位国家: 美国
• 财政年份: 2011
• 资助国家: 美国
• 起止时间: 2011-09-20 至 2016-08-31
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/8708556
• 关键词: Address; Biological; Biological Assay; Cancer Patient; Categories; Classification; Computer software; Copy Number Polymorphism; DNA Sequence; DNA copy number; Data; Data Element; Data Set; Databases; Development; Disease; Drug Formulations; Elements; Gene Expression; Gene Expression Profile; Generations; Genes; Genomics; Glioblastoma; Goals; Graph; Individual; Joints; Lead; Learning; Left; Malignant Neoplasms; Malignant neoplasm of lung; Measurement; Medicine; Methodology; Methods; Methylation; Mining; Modeling; Network-based; Noise; Normal Statistical Distribution; Normal tissue morphology; Other Genetics; Outcome; Pathway interactions; Patients; Pattern; Performance; Principal Component Analysis; RNA Sequences; Research; Sampling; Signal Transduction; Single Nucleotide Polymorphism; Statistical Methods; Statistical Models; Stem cells; Structure; Techniques; Technology; Testing; Time; Tumor Subtype; Validation; Validity of Results; base; cancer therapy; cancer type; clinically relevant; disorder subtype; flexibility; gene discovery; leiomyosarcoma; malignant breast neoplasm; novel strategies; outcome forecast; response; stem cell biology; stem cell population; tool; tumor

PROJECT SUMMARY. This project involves the development of statistical methodology for the analysis of large- scale genomic data, such as gene expression, DNA copy number, and DNA sequencing data. In genomic studies, the goal is often to identify signal in the data in an unsupervised way. For instance, given the gene expression measurements for a set of patients with lung cancer, one might wish to discover previously unknown lung cancer subtypes that are characterized by distinct gene expression signatures and that might differ with respect to prognosis or response to therapy. However, the search for signal in genomic data is made difficult by the fact that the number of variables (e.g. genes) is generally orders of magnitude greater than the number of observations (e.g. lung cancer patients). As a result, principled methods must be developed to discover signal without overfitting. Furthermore, there is a need for objective ways to assess the validity of results obtained. This proposal has four specific aims, each of which involves the development of a new statistical method for solving a problem that arises in the analysis of genomic data. Aim 1: A method to learn multiple related genomic networks at once. For instance, one might expect that the gene expression networks for cancer and normal tissues will look similar to each other, with certain specific differences. The current proposal will provide a way to learn both networks simultaneously, in order to identify gene pathways that are perturbed in cancer. The proposed approach involves applying shrinkage penalties to the Gaussian graphical model formulation for network estimation. Aim 2: A principled approach for simultaneously clustering the rows and columns of a data matrix (e.g. patients and genes). The standard approach for discovering signal in genomic data involves clustering rows and columns independently, but the proposed approach will have increased power to discover biologically relevant clusters. The proposed approach involves applying shrinkage penalties to the matrix-variate normal distribution. Aim 3: A tool for the integrative analysis of multiple genomic data types collected on a single set of patient samples. For instance, if gene expression data, copy number data, and methylation data are collected for a single set of samples, then this will allow for the discovery of subsets of patients that are characterized by particular signatures of gene expression, copy number variation, and methylation. This could lead to the discovery of clinically relevant subtypes of cancer and other diseases. The proposed approach is an extension of the approach described in Aim 2. Aim 4: A flexible framework for the validation of clusters discovered in structured genomic data, such as DNA copy number and single nucleotide polymorphism data, in order to determine whether clusters discovered reflect signal or simply noise. The proposed approach is related to cross-validation, and will be extended to develop a method for the validation of other unsupervised statistical tools, such as those described in Aims 1-3 above. The statistical tools that result from the proposed research will be implemented in freely available software.

项目摘要。该项目涉及开发统计方法，用于分析大型 比例基因组数据，例如基因表达，DNA拷贝数和DNA测序数据。在基因组研究中， 该目标通常是以无监督的方式识别数据中的信号。例如，给定基因表达 一组肺癌患者的测量，人们可能希望发现以前未知的肺癌 以独特的基因表达特征的亚型，在预后方面可能有所不同 或对治疗的反应。但是，基因组数据中寻找信号的搜索很难通过数字 变量（例如基因）的数量级通常大于观察次数（例如肺癌 患者）。结果，必须开发有原则的方法以发现信号而不过度拟合。此外， 需要客观的方法来评估所获得的结果的有效性。 该提议具有四个具体目标，每个目标都涉及开发一种新的统计方法 解决基因组数据分析中出现的问题。目标1：一种学习多个相关基因组的方法 一次网络。例如，人们可能期望癌症和正常组织的基因表达网络 看起来彼此相似，并且存在某些具体差异。当前的建议将提供一种学习两者的方法 网络同时为了鉴定癌症干扰的基因途径。提出的方法 涉及将收缩惩罚应用于高斯图形模型公式以进行网络估计。目标2： 同时聚集数据矩阵的行和列的原则方法（例如，患者和基因）。 在基因组数据中发现信号的标准方法涉及独立聚类行和列， 但是拟议的方法将增加发现与生物学相关簇的能力。提议 方法涉及将收缩惩罚应用于矩阵变化的正态分布。目标3：一种工具 对单个患者样本收集的多种基因组数据类型的综合分析。例如，如果 为一组样品收集基因表达数据，拷贝数数据和甲基化数据，然后收集 将允许发现以基因表达的特定特征的患者的子集， 拷贝数变化和甲基化。这可能导致发现癌症的临床相关亚型 和其他疾病。提出的方法是AIM 2中描述的方法的扩展。目标4：灵活 验证在结构化基因组数据中发现的簇的框架，例如DNA拷贝数和单个 核苷酸多态性数据，以确定簇是否发现了信号还是简单的噪声。这 建议的方法与交叉验证有关，并将扩展以开发一种验证其他方法的方法 无监督的统计工具，例如上面的目标1-3中描述的工具。 拟议研究产生的统计工具将在免费提供的软件中实施。