III-CXT: Learning from graph-structured data: new algorithms for modeling physical interactions in cellular networks

从图结构数据中学习：模拟细胞网络中物理相互作用的新算法细胞的复杂行为源于成千上万个基因及其产物的分子相互作用的复杂网络。了解这个网络如何运作并预测其行为是生物学的主要目标，对生命科学、医学和生物技术有着广泛的影响。过去十年的基因组信息革命为研究蜂窝网络提供了新的系统级和数据驱动的方法。特别是，利用机器学习来模拟基因调控网络——通过与非编码DNA结合的调控蛋白来开启和关闭基因——已经成为系统生物学的一个核心问题。现在，用于测量蛋白质之间以及蛋白质与DNA之间物理相互作用的新型高通量技术的爆炸式增长为基因调控的计算建模提供了新的数据集成挑战。这些新数据都可以被看作是图形结构的数据，或者是物理交互网络。这个项目的核心计算目标是开发新的机器学习算法来利用图结构数据，包括：(1)利用高效的图挖掘进行提升；(2)基于子图直方图的图核；(3)基于信息的图划分。这些新算法将用于将物理相互作用网络数据整合到基因调控模型中，以便更好地代表潜在的生物机制。重点将是两个基本的建模问题：推断信号转导途径和在DNA序列和相互作用的调节蛋白水平上建模顺式调节模块。这些算法将被应用于公开可用的数据和由一名研究人员提供的主要基因表达数据，以研究酵母中的缺氧和哺乳动物神经细胞对环境毒素的反应。该项目将学习系统级模型，从而对基因调控的潜在机制产生新的见解，并为更广泛的生物学发现开辟道路。所有数据、结果和源代码将通过网络（http://www.cs.columbia.edu/ compbio/cellular-networks）公开，并通过课程和生物信息学软件包传播。该项目还将为本科生提供干湿联合实验室项目和推广活动的研究机会，向纽约市公立高中学生介绍新的跨学科科学领域。

III-CXT: Learning from graph-structured data: new algorithms for modeling physical interactions in cellular networksThe complex behavior of the cell derives from an intricate network of molecular interactions of thousands of genes and their products. Understanding how this network operates and predicting its behavior are primary goals of biology and have broad implications for life science, medicine and biotechnology.The genomic information revolution of the last ten years has enabled new systems-level and data-driven approaches for studying cellular networks. In particular, using machine learning to model gene regulatory networks---the switching on and off of genes by regulatory proteins that bind to non-coding DNA---has emerged as a central problem in systems biology. Now, an explosion of new high-throughput technologies for measuring physical interactions between proteins and between protein and DNA provides a new data integration challenge for computational modeling of gene regulation. These new data can all be viewed as graph-structured data, or physical interaction networks.The central computational goal of this project is to develop new machine learning learning algorithms for exploiting graph-structured data, including: (1) boosting with efficient graph mining; (2) graph kernels based on subgraph histogramming; and (3) information-based graph partitioning. These new algorithms will be used to integrate physical interaction network data into models of gene regulation in order to better represent underlying biological mechanisms. The focus will be two fundamental modeling problems: inferring signal transduction pathways and modeling cis regulatory modules at the level of DNA sequence and interacting regulatory proteins. The algorithms will be applied both to publicly available data and to primary gene expression data provided by one of the investigators to study the hypoxia in yeast and the response to environmental toxins in mammalian neural cells.This project will learn systems-level models that lead to new insight into the underlying mechanisms of gene regulation and open the way to broader biological discoveries. All data, results and source code will be publicly available via the Web (http://www.cs.columbia.edu/ compbio/cellular-networks) and disseminated through courses and bioinformatics software packages. The project will also create undergraduate research opportunities for joint dry and wet lab projects and outreach activities to introduce New York City public high school students to new interdisciplinary areas of science.