CAREER: A Data-Driven Network Inference Framework for Context-Conditioned Protein Interaction Graphs

生物系统可以用图形来研究，其中节点代表实体(例如，蛋白质)，边代表相互作用(例如，物理结合、功能依赖)。重要蛋白质相互作用网络的识别使人们能够对生命、进化变化、疾病研究和药物开发的原理有新的见解。蛋白质相互作用图的网络连接和功能由上下文决定：遗传、环境和药物等小分子。然而，由于生物技术对图形数据收集的限制，到目前为止，几乎所有的蛋白质相互作用网络都是在单一的静态条件下进行检查的。因此，该方案的研究目标是设计新颖高效的机器学习算法来识别特定上下文的蛋白质相互作用图。识别特定背景的蛋白质网络具有重要的社会意义的生物医学应用，例如研究多个细胞阶段的细胞发育，或在白血病背景下研究不同药物治疗下的细胞变化。通过与公共卫生基因组学中心和弗吉尼亚大学医学院艾米丽·库里克癌症中心的合作，这两个应用程序将作为该项目的评估组成部分进行探索。这项拟议的研究预计也会影响其他领域，例如，社会网络发现和大脑连接的特定条件网络推理。拟议的职业生涯计划将产生以研究的跨学科性质为基础的教育和外联举措。这些计划包括：(A)设计处理现实网络推理问题和数据的新课程项目；(B)开发新的教学技术，以最先进的结构学习问题为样本项目，培训研究生的专业技能，如“如何教学”或“如何做研究”；(C)通过大学本科生顶尖项目，让本科生参与网络学习研究；(D)通过在有高中生参加的大学工程导论(ITE)计划上的演讲，提高K-12年级学生对图形学习研究的意识；以及(E)加强与弗吉尼亚大学医学院社区的互动，特别是通过公开发布和对该项目创建的计算工具进行教程。过去十年见证了基因组技术的革命，能够同时测量数千个分子实体(例如，基因或蛋白质)。下一代测序技术产生的大量全基因组数据提供了前所未有的覆盖范围，涵盖了相关基因产品的大规模、有上下文条件的签名，这些签名具有在每种上下文中推断网络连接和功能的巨大潜力。该提案将开发一套新颖的机器学习方法，用于从高维、异质和噪声的多上下文分子签名数据集中推断特定上下文的网络。为了克服这些数据挑战，本研究包括以下三个相关任务：(I)开发新的可扩展的结构学习算法，从多个不同条件下聚集的数据样本中估计多个不同但相关的稀疏高斯图形模型(SGM)；(Ii)在多任务稀疏高斯图形模型的框架内，为建模和检测模块(即多蛋白质组)开发新的学习策略；(Iii)将上述结构学习模型扩展到非高斯情况、考虑部分观察网络的半监督环境和有监督的疾病诊断环境。有关该项目的其他信息，包括出版物、算法的开源实现、数据集和教育材料，将通过该项目的网站共享：http://www.cs.virginia.edu/yanjun/context_graph/

Biological systems can be studied as graphs, where nodes represent entities (e.g., proteins) and edges represent interactions (e.g., physical binding, functional dependency). The identification of important protein interaction networks enables new insights into principles of life, evolution change, disease study, and drug development. The network wiring and function of a protein interaction graph is determined by context: genetics, environment, and small molecules such as drugs. However, almost all protein interaction networks, to date, have been examined under a single static condition, due to limitations of biotechnologies for graph data collection. Therefore, the research objective of this proposal is to design novel and efficient machine-learning algorithms to identify context-specific protein interaction graphs. Identifying context-specific protein networks has biomedical applications of social importance, such as studying cellular developments across multiple cell stages or investigating cellular changes with different drug treatments in the context of leukemia. Both applications will be explored as evaluation components of the project through collaborating with the Center for Public Health Genomics and the Emily Couric Cancer Center at UVA School of Medicine. The proposed research is expected to impact other domains as well, for instance, social-network discovery and condition-specific network inference for brain connectivity. The proposed career plan will result in educational and outreach initiatives that build on the interdisciplinary nature of the research. These plans include: (a) designing new course projects that work on real-life network-inference problems and data; (b) developing novel instructional techniques to train graduate students professional skills such as "how to teach'' or "how to do research" using state-of-the-art structural learning problems as sample projects; (c) involving undergraduates in network learning research through UVA undergraduate capstone projects; (d) increasing awareness of graph-learning research among K-12 students through presentations at the UVA Introduction to Engineering (ITE) Program involving high school students; and (e) enhancing interactions with the UVA Medical School Community, especially through public release and tutorials of computational tools created from this project.The past decade has seen a revolution in genomic technologies that enable the simultaneous measurement of thousands of molecular entities (e.g., genes or proteins). The flood of genome-wide data generated by next-generation sequencing technologies has provided an unprecedented coverage of large-scale, context-conditioned signatures of relevant gene products that have great potential to infer network connectivity and function in each context. The proposal will develop a suite of novel machine-learning methods for inference of context-specific networks from multi-context molecular signature datasets that are high dimensional, heterogeneous and noisy. Aiming to overcome these data challenges, the proposed research includes the following three related tasks: (i) develop new and scalable structural learning algorithms to estimate multiple different but related sparse Gaussian Graphical Models (sGGMs) from data samples aggregated across multiple distinct conditions, (ii) develop novel learning strategies for modeling and detecting modules (i.e., multi-protein groups) within the framework of multitasking sGGMs, (iii) extend the above structural learning models to non-Gaussian cases, semi-supervised settings considering partial-observed networks and supervised disease diagnosis settings. Additional information about the project, including the publications, open-source implementations of algorithms, data sets and educational materials will be shared through the project website: http://www.cs.virginia.edu/yanjun/context_graph/