Novel methods for integrative computational biology

目前生物医学研究面临的挑战包括信息过载：需要结合大量的结构化和非结构化信息，以及需要识别、表征、优化、链接、验证和解释复杂数据中的模式。大多数生物数据和生物实体之间的关系被方便地建模为图表。图形数据库为集成不同的实体提供了一个高性能的解决方案，并为机器学习和推理提供了一个可扩展的平台。数据挖掘、图论和本体论是成功学习系统的重要组成部分。这些类型图的质量、覆盖率和注释为模型和假设的生成提供了必要的平台。 我们建议开发一个可扩展的分布式体系结构来存储、分析集成的图形数据，以支持高级机器学习和数据建模。该资源将提供可用蛋白质组、联盟和图谱实体的全面覆盖，以建立实验检测的、同源的和预测的组织和条件特定蛋白质相互作用的数据库。我们将改进和扩展我们的关联挖掘算法FpClass，这是一个用于相互作用蛋白质之间加权链接预测的机器学习系统。我们将使用图灵完全和性能优化的图形查询语言Gremlin将该方法扩展到在序列、组织和特定条件的特定上下文中的集成图形模型下预测基因和蛋白质之间的各种类型的链接。 本体论将用于支持多个视点和上下文。图论和数据库将提供必要的基础设施和复杂的模式推理平台。机器学习和数据挖掘将提供新的模型，而概率建模将支持处理不完整、矛盾和模糊的信息。我们将使用导航器来支持可视化数据挖掘和类型化图形的交互探索。 我们将使用多个公开可用的数据集来测试和验证开发的平台。这项研究将产生新的算法，经过验证后，它们的应用将导致在分子水平上更好地理解复杂疾病。算法、结果和相关数据将公之于众，以鼓励在这个日益重要的领域进行进一步的计算研究。与深度学习一起，这些方法正在给应用领域带来革命性的变化，因为这些系统的性能经常超过领域专家的能力和生物检测灵敏度和错误发现率。 所提出的研究将推进计算方法及其在高通量系统生物学应用中的适用性。这项建议的一个重要功能是培训生物信息学专业人员，这方面的需求仍然很大。

Current challenges in biomedical research include information overload: the need to combine vast amounts of structured and unstructured information, and the need to identify, characterize, optimize, link, validate and interpret patterns in complex data. The majority of biological data and relationships between biological entities is conveniently modeled as a graph. Graph databases provide a high-performance solution to integrating diverse entities and providing a scalable platform for machine learning and inference. Data mining, graph theory and ontologies are essential components of successful learning systems. Quality, coverage and annotation of these typed graphs provide the necessary platform for model and hypotheses generation. We propose to develop a scalable, distributed architecture to store, analyze integrated graph data to support advance machine learning and data modeling. This resource will provide comprehensive coverage of available proteome, federate and map entities to a database of experimentally detected, orthologous, and predicted tissue- and condition-specific protein interactions. We will improve and extended our association mining algorithm, FpClass, a machine learning system for weighted link prediction between interacting proteins. We will extend the method to predict links of various types between genes and proteins under the integrated graphical model in sequence, tissue, and condition-specific contexts using the Turing Complete and performance-optimized graph query language Gremlin. Ontologies will be used to support multiple viewpoints and contexts. Graph theory and databases will provide necessary infrastructure and complex pattern inference platform. Machine learning and data mining will contribute new models, while probabilistic modeling will support handling incomplete, contradictory and ambiguous information. We will use NAViGaTOR to support visual data mining and interactive exploration of the typed graphs. We will test and validate developed platforms using multiple, publicly available datasets. This research will generate novel algorithms, and after validation, their application will lead to improved understanding of complex diseases on a molecular level. Algorithms, results and relevant data will be made publicly available to encourage further computational research in this increasingly important area. Together with deep learning, these approaches are revolutionizing application areas since the performance of these systems frequently exceed domain experts' abilities and biological assay sensitivity and false discovery rates. The proposed research will advance computational approaches and their applicability to high-throughput systems biology applications. An important function of this proposal is the training of bioinformatics professionals for which there is still great unfulfilled demand.