ATD: Collaborative Research: Statistical Ensembles for the Identification of Bacterial Genomes

这项研究的重点是通过以以前从未研究过的方式整合基因组学和统计学的工具来减少生物恐怖主义威胁。研究人员开发了新的统计理论和计算工具，以基于下一代测序数据进行准确的病原体检测。主要研究方向包括(I)通过序列丰富进行分类；(Ii)经验聚类和参考基因组的比较；以及(Iii)分层对数线性模型中的收缩估计和模型选择。除了深入描述这些新的统计推断技术的理论特性外，研究人员还对它们在细菌基因组检测和鉴定方面的实际重要性进行了彻底的评估。这项评估使用了来自人类微生物组项目、NCBI短文档案馆、欧洲生物信息学研究所和博德研究所等来源的公开数据。这一新方法的适用性很广，涉及到很难选择适当类别的候选统计模型的高维环境。研究人员研究了统计集合，即已被证明比任何单一统计方法提供更可靠推断的技术组合。与现有的将同一类的模型组合在一起的工作不同，这个新的框架涉及跨越类边界的集成，并以最佳方式组合来自几个模型类的多个模型的推理。与现有方法相比，这些集成有望具有明显的优势，例如对模型错误指定的稳健性和改进的预测性能。在这项建议中开发的新统计方法有可能通过快速识别微生物基因组的差异并将其准确归类为无害或潜在致病的微生物，大幅改善联邦和国际机构对生物恐怖主义袭击的反应。这些病原体检测算法对信息技术和民用基础设施的影响是最大限度的，方法是采用方便用户的、开放源码的计算工具和软件，向公众免费提供。该项目还为学生和博士后研究员提供了重要的教育和指导内容，他们有兴趣提高我们迅速和适当地应对(I)生物恐怖主义事件和(Ii)对公共卫生的微生物威胁的能力。

This research focuses on reducing bioterrorism threat by integrating tools from genomics and statistics in ways that have not been previously examined. The investigators develop novel statistical theory and computational tools for accurate pathogen detection based on next generation sequencing data. Key research directions involve (i) classification by sequence enrichment; (ii) comparison of empirical clusterings and reference genomes; and (iii) shrinkage estimation and model selection in hierarchical log-linear models. In addition to an in-depth characterization of the theoretical properties of these new statistical inference techniques, the investigators perform a thorough assessment of their practical importance in the context of the detection and identification of bacterial genomes. This assessment is done using publicly available data from sources such as the Human Microbiome Project, the NCBI Short Read Archive, the European Bioinformatics Institute, and the Broad Institute. The applicability of this new methodology is broad and relates to high-dimensional settings in which choosing an appropriate class of candidate statistical models is difficult. The investigators study statistical ensembles, combinations of techniques that have been shown to provide more reliable inferences than any single statistical approach. As opposed to existing work which combines models from the same class, this new framework concerns ensembles that cross class boundaries and optimally combine inferences from multiple models from several model classes. These ensembles are expected to have distinct advantages over existing approaches, such as robustness to model misspecification and improved predictive performance.The new statistical methodology developed in this proposal has the potential to substantially improve the response of federal and international agencies to a bioterrorism attack through a rapid identification of differences in microbial genomes and their accurate classification as harmless or potentially pathogenic. The impact of these algorithms for pathogen detection on both information technology and civil infrastructure is maximized through their implementation in user-friendly, open-source computational tools and software that will be freely available to the public. The project also has a significant educational and mentorship component for students and postdoctoral fellows who are interested in enhancing our ability to respond rapidly and appropriately to (i) incidents of bioterrorism, and (ii) microbial threats to public health.