EAGER: Transcript-Based Differential Expression Analysis for Population Data Without Predefined Conditions

随着精准医学的出现，对更灵敏的分子生物标志物的需求增加。分子生物标记物发现的一个基本计算步骤是识别在不同样本中不同表达的基因。本项目研究新的算法方法，用于在转录水平上对没有预先定义的生物条件的样本进行差异表达分析。这样的分析对于对人群(或队列)数据的临床和生物学研究都是至关重要的。例如，它可以用来发现分子生物标记物，将癌症样本分类为亚型，以便为每一亚型开发更好的诊断和治疗方法。它还可以用来描述参与不同生物过程的单个细胞的特征。基于本项目提出的新的分析方法，可以构建高效的软件工具，帮助生物学家发现比现有方法更敏感的生物标志物。具体地说，本项目研究了三种对种群数据进行差异转录表达分析的方法。前两种方法将基因的外显子或完整转录本作为基本表达元件，然后对这些表达元件应用基因水平的差异表达分析方法。第三种方法是前两种方法的混合体。它使用剪接图来表示基因的转录本，并使用一种新的模分解算法将图划分为对应于独立可选剪接事件的小组件。此外，还使用了一种稳健的聚类算法来处理输入总体中的任意数量的条件。这三种方法的软件实现都在模拟和真实序列数据上进行了广泛的校准和测试，以确定它们对公众的实用价值。

With the emergence of precision medicine, there is increased demand for more sensitive molecular biomarkers. A fundamental computational step in the discovery of molecular biomarkers is to identify genes that are expressed differently across different samples. This project investigates new algorithmic approaches for performing differential expression analysis at the transcript level for samples without predefined biological conditions. Such an analysis is critical to both clinical and biological studies on population (or cohort) data. For example, it can be used to discover molecular biomarkers to classify cancer samples into subtypes so that better diagnosis and therapy methods can be developed for each subtype. It can also be used to characterize individual cells involved in different biological processes. Efficient software tools can be built based on the new analysis approaches proposed in this project which can help biologists to discover more sensitive biomarkers than the existing methods. Specifically, this project studies three approaches for differential transcript expression analysis on population data. The first two approaches treat either the exons or full transcripts of a gene as the basic expression elements and then apply a gene-level differential expression analysis method on these expression elements. The third approach is a hybrid of the first two. It uses a splice graph to represent the transcripts of a gene and a new modular decomposition algorithm to partition the graph into small components that correspond to independent alternative splicing events. Moreover, a robust clustering algorithm is employed to deal with an arbitrary number of conditions in the input population. The software implementations of the three approaches are calibrated and tested extensively on both simulated and real sequence data to establish their practical utility to the public.