Recovering reproducible and local signal in genomic data

挑战。当今生物科学中最重要的挑战之一是阐明复杂的 测量数十万个变量的实验可以被分析，以在以下情况下产生一致的全局信号 重复，以识别与组织、癌症类型或种群结构相关的局部信号。 重要的是，我们必须包括不同研究和技术混杂因素控制的内在多样性，如 这是这项任务的一部分。 来自高维生物实验的大多数测量结果都显示出由生物来源引起的差异，例如 作为属于大脑中不同组织或不同位置的基因。而有些组件则会在多个 在组织中，全局生物信号比虚假信号更有可能在多个组织中重复存在。我们的挑战是 系统和可靠地识别全球生物因素，并估计每项研究特有的信号。 目标。为了迎接这一挑战，我们提出了一种新的概念，它结合了元分析和统计学的思想 建模降维。我们假设，人们可以开发高维数据约简技术，同时 时间作为多学习工具来提取一致信号和局部特定信号。这一建议发展了统计方法 用于在多项癌症研究中识别共享的和研究特定的信号。在这项工作中，理解共享的 信号--这里指的是不同癌症类型之间共享的基因共表达--以及每个研究特有的信号。这项提议将 通过建立多元Logistic回归和因子分析方法的新类别来试验这一概念。关键是要分解 每项研究的数据都涉及潜在的维度，其中一些是全球性的，而另一些则不是，而且只针对局部信号。 这将同时实现两个目标：了解研究中共享的可重复生物特征，以及识别 每项研究的具体变化。具体目标包括方法设计、软件开发和应用。 冲击力。这项研究产生的概念、方法和软件工具将直接影响到 生物医学社区在多个高通量生物学研究中重复识别稳定的信号并捕获 本地信号。我们的工具还将实现更可靠的人工产物识别，从而促进更有效的实验 设计和指导技术发展。我们还希望对数据科学产生超越基因组学的影响。我们的研究将是 第一次有机会评估共享潜在因素以及估计局部潜在结构的新概念。这个 拟议的工作可随后为将这一概念扩展到 多种另一种降维和机器学习技术。

Challenge. One of the most important challenge in biological science today is to elucidate the extent to which complex experiments, which measure hundreds of thousands of variables, can be analyzed to generate consistent and global signal when repeated, to identify local signal related to tissues, cancer types or population structure. Importantly, we must include the intrinsic diversity of variation across different studies and control for technical confounders as part of this task. Most measurements from high-dimensional biological experiments display variation arising both from biological sources, such as genes belonging to a different tissue or different positions in the brains. While some components reappear across multiple tissues, global biological signal is more likely than spurious signal to be reproducibly present in multiple tissues. Our challenge is to systematically and reliably identify the global biological factors, and estimate the signal specific to each study. Aims. In order to meet this challenge, we propose a novel concept that combines ideas from meta-analysis and statistical modeling dimension reduction. We posit that one can develop high-dimensional data reduction techniques that at the same time function as multi-study tools to extract consistent signal and local specific signal.This proposal develops statistical methods for identifying shared and study-specific signal across multiple cancer studies. In this work, it is crucial to understand the shared signal - here, gene co-expression shared across different cancer types - and the signal specific to each study. This proposal will pilot this concept by building novel classes of multi-logistic regression and factor analysis methods. The key is to decompose data from each study into latent dimensions, some of which are global while some are not and only specific to a local signal. This will simultaneously achieve two goals: learning reproducible biological features shared among studies, and identifying the variation specific of each study. Specific aims include methodology design, software development and applications. Impact. The concepts, approaches, and software tools generated by this research will have a direct impact on the ability of the biomedical community to reproducibly identify stable signals across multiple high-throughput biology studies and to capture local signals. Our tools will also enable a more reliable identification of artifacts and thus facilitate more efficient experimental designs and guide technological development. We also hope to impact data sciences beyond genomics. Our study will be the first opportunity to evaluate the novel concept of sharing latent factors as well as estimating local latent structures. The proposed work could subsequently provide the inspiration, as well and the practical foundation, for expanding this concept to a variety of another dimension reduction and machine learning techniques.