CAREER: Learning the Chromatin Network from ChIP-Seq Data

生物体的每个细胞共享相同的核DNA序列(基因组)，但不同类型的细胞启动不同的基因集来执行其独特的功能。要正确地打开和关闭所需的基因集，需要数百个调控分子的协调行动。调节器与基因组相互作用：了解这种情况是如何发生的，是生物学研究人员试图回答的最重要的问题之一。该项目将使用染色质免疫沉淀测序(ChIP-Seq)数据，该数据测量特定调控因子在基因组上的位置：这可能是基因组中数十亿个可能的位置中的零个、一个或几百个位置。监管机构经常一起行动，一个监管机构可能会改变另一个监管机构的行动，因此拟议工作的另一部分是发现监管机构共同本地化的地点。然而，共同本地化本身并不能表明监管机构是直接互动，还是通过中间人间接互动。该项目的总体目标是开发一个计算框架，利用数千个芯片序列数据集确定监管机构之间的直接和间接相互作用。这一框架将开发新的统计和机器学习技术，以克服现有方法的局限性。开发的技术将被应用于回答以下基本问题：数百个调节器如何相互作用来调节基因组？这些调节因子的相互作用在不同类型的细胞中有何不同？这些相互作用在不同物种(人、老鼠、苍蝇和蠕虫)之间有何不同？新方法的实施将公开，以帮助许多其他科学家研究人类基因组是如何工作的。这个项目本质上是跨学科的，并通过项目课程和推广活动非常重视跨学科教育。识别染色质调节因子之间的相互作用，如转录因子和组蛋白修饰，对于理解基因组调控至关重要。为了推断这种相互作用的网络，这项研究将比较多个染色质免疫沉淀-测序(CHIP-SEQ)数据集，每个数据集都测量染色质调节因子的全基因组定位。共本地化可能表明两个调节者通过传递性相互作用直接或间接地相互作用。为了识别直接相互作用，该提议旨在开发新的网络推理方法来推断大量芯片序列数据集之间的条件依赖关系(即，不通过网络中的任何其他变量来解释的相关性)。虽然网络推理已成为其他类型数据的常用分析工具，如基因表达数据，但芯片序列数据集的巨大规模和数据中存在的强烈冗余限制了现有网络推理方法的使用。为了解决这些挑战，本研究提出了一个新颖的机器学习(ML)框架，以实现从大量芯片序列数据中进行网络推理：1)基于包含冗余的整个ENCODE芯片序列数据来推断染色质网络的高效ML方法；2)通过结合其他类型的基因组数据来联合推断特定于上下文的染色质网络和相关基因组上下文的新ML方法；以及3)新的ML方法来学习跨物种的保守染色质网络并预测染色质因子的相互作用，即使这些因素没有在所研究的物种中测量到。欲了解更多信息，请访问项目网站：http://suinlee.cs.washington.edu/projects/chromnet.

Each cell of an organism shares the same nuclear DNA sequence (genome), but different types of cells turn on different sets of genes to carry out their unique functions. To correctly turn on and off the needed sets of genes requires coordinated action by several hundred regulatory molecules. The regulators interact with the genome and with each other: understanding how this happens is one of the most important questions biological researchers are trying to answer. This project will use chromatin immunoprecipitation-sequencing (ChIP-Seq) data, which measures where a particular regulator is located on the genome: this may be zero, one or a few hundred positions out of the billions possible in a genome. Regulators often act together and one may change the action of another, so another part of the proposed work is to discover sites where regulators are co-localized. However, co-localization alone does not show whether regulators interact directly, or indirectly through an intermediate. The overall goal of this project is to develop a computational framework that identifies direct and indirect interactions among regulators, using thousands of ChIP-seq datasets. This framework will develop novel statistical and machine learning techniques to overcome the limitations of existing methods. The techniques that are developed will be applied to answer the following fundamental questions: How do hundreds of regulators interact with each other to regulate the genome? How do these regulator interactions differ across cell types? How do these interactions differ across different species (human, mouse, fly and worm)? The implementation of the new methods will be made publicly available to help many other scientists to study how the human genome works. This project is interdisciplinary in nature and has significant emphasis on interdisciplinary education, through project courses and outreach activities.Identifying the interactions among chromatin regulators, such as transcription factors and histone modifications, is of paramount importance to understand genome regulation. To infer this network of interactions, this research will compare multiple chromatin immunoprecipitation-sequencing (ChIP-seq) datasets, each measuring genome-wide localization of a chromatin regulator. Co-localization may indicate that two regulators interact directly or indirectly through transitive interactions. To identify direct interactions, the proposal aims to develop novel network inference methods to infer conditional dependence relationships (i.e., correlation not explained via any other variables in the network) among a large number of ChIP-seq datasets. While network inference has become a commonly used analysis tool for other types of data, such as gene expression data, the immense size of the ChIP-seq data sets and the strong redundancies present in the data limit the use of existing network inference methods. To resolve these challenges, this research proposes a novel machine learning (ML) framework to enable network inference from large collections of ChIP-seq data: 1) efficient ML methods to infer the chromatin network based on the entire ENCODE ChIP-seq data that contain redundancies; 2) new ML methods to jointly infer the context-specific chromatin networks and the associated genomic contexts by incorporating other types of genomic data; and 3) new ML methods to learn a conserved chromatin network across species and predict chromatin factor interactions even if the factors are not measured in the species of study. For further information see the project web site at: http://suinlee.cs.washington.edu/projects/chromnet.