CAREER: A Comprehensive and Lightweight Framework for Transcriptome Analysis

在过去的十年里，测序技术已经发展起来，能够对各种生物体和组织类型的基因表达进行分析。这些技术允许在转录组范围内研究基因表达在不同条件、不同刺激下和不同疾病状态下的变化。这些技术在推动基础科学(如了解细胞生物学)和应用科学(如药物开发方法)方面具有变革性作用。然而，这些技术产生的海量数据带来了一系列计算挑战，例如发现样本是否包含以前没有注释的基因，准确确定这些基因的序列，以及量化样本中表达的所有基因的丰度。人们一直致力于开发可靠的计算方法来处理这些数据。然而，就精度而言，即使是最好的现有解决方案有时也不能令人满意，而且考虑到新数据产生的速度如此之快，这些解决方案在计算上正变得繁重。该项目的目标是开发新一代准确和轻量级的方法，利用测序数据分析基因和转录本的表达。这些工具将应用新的数据结构和算法思想来解决绘制测序读数图、发现和组装新的转录本以及准确而有力地量化基因表达的问题。此外，这些方法将在现有技术和新出现的协议的背景下工作，这些协议允许测量数千个单个细胞的细胞特异性基因表达。该项目产生的方法和软件将通过比现有方法更灵敏和准确来帮助实现新的发现，通过减少计算需求来降低成本，并将通过比现有方法更快地产生结果来加快分析速度。该项目的推广目标包括创建教育媒体，包括视频和播客系列，帮助向执业生物学家以及对科学感兴趣的公众传达新的计算基因组学方法的关键见解和好处。轻量级量化方法简化了许多常见的转录分析，如对注释良好的生物体和常见组织类型的差异表达测试。然而，仍然存在大量的挑战，阻碍了轻量级方法在许多分析任务中的使用，例如，当应该考虑新的转录本时，或者当内含子保留等事件可能发挥重要作用时。这项工作将提高轻量级转录组分析方法的准确性和基本能力。具体地说，将开发一种新的基于图形的数据结构，用于索引参考序列的集合。围绕这一索引将建立一个轻量级的比对工具，该工具将纳入一个统计模型，该模型允许在大型样本集合中共享剪接上下文，以指导和通知困难的比对问题。还将根据对转录序列及其丰度的联合可能性进行建模的新方法，开发一种用于联合转录发现和量化的多样本方法。高效的似然因式分解将允许这种方法在计算上保持方便。最后，将开发一套用于处理和量化高通量单细胞RNA-SEQ数据的工具。这些工具将采用一种新的方法来联合解决细胞条形码、UMI重复数据删除和基因表达估计问题，并在统一的统计框架中进行。基本模型将在细胞之间共享统计信息，以改进聚类和量化，并以数据支持的分辨率分析表达，即作为可区分的异构体的组。所有这些工具都将作为高质量的开源软件发布。

Over the past decade, sequencing technologies have been developed that enable the profiling of gene expression across a wide variety of organisms and tissue types. These technologies allow the investigation, on a transcriptome-wide scale, of how gene expression changes in different conditions, under various stimuli, and in different disease states. These technologies are transformative in progressing basic science (e.g., understanding cell biology) and applied science (e.g., approaches to drug development). However, the deluge of data produced by these technologies brings with it a host of computational challenges, such as discovering if samples contain genes previously not annotated, accurately determining the sequence of these genes, and quantifying the abundance of all the genes expressed in a sample. Much effort has been dedicated to developing reliable computational methods for processing this data. Yet, even the best existing solutions are sometimes unsatisfactory in terms of their accuracy, and are becoming computationally burdensome given the rapid rate at which new data is being produced. The goal of this project is to develop a new generation of accurate and lightweight methods for analyzing gene and transcript expression using sequencing data. These tools will apply new data structures and algorithmic ideas to the problems of mapping sequencing reads, discovering and assembling new transcripts, and accurately and robustly quantifying gene expression. Further, these methods will work in the context of both established technologies and the newly-emerging protocols that allow measuring cell-specific gene expression across thousands of individual cells. The methods and software produced as a result of this project will help enable new discoveries by being more sensitive and accurate than existing approaches, will reduce costs by decreasing computational demands, and will speed up analyses by producing results more quickly than existing approaches. The outreach goals of this project include the creation of educational media including videos and a podcast series that will help convey key insights and benefits of new computational genomics methods to both practicing biologists as well as to the scientifically-interested public at large.Lightweight quantification methods streamline many common transcriptomic analyses, like differential expression testing in well-annotated organisms and common tissue types. Yet, substantial challenges remain that prevent the use of lightweight methods in many analysis tasks, e.g., when novel transcripts should be considered, or when events such as intron retention may play an important role. This work will advance the accuracy and fundamental capabilities of lightweight transcriptome analysis methods. Specifically, a new graph-based data structure will be developed for indexing a collection of reference sequences. A lightweight alignment tool will be built around this index that will incorporate a statistical model that allows sharing of splicing contexts across large collections of samples to guide and inform difficult alignment problems. A multi-sample methodology for joint transcript discovery and quantification will also be developed, based on new approaches to modeling the joint likelihood of transcript sequences and their abundances. Efficient likelihood factorizations will allow this approach to remain computationally convenient. Finally, a suite of tools for processing and quantifying high-throughput, single-cell RNA-seq data will be developed. These tools will adopt a novel approach for solving the cell barcoding, UMI deduplication, and gene expression estimation problems jointly, and in a unified statistical framework. The underlying model will share statistical information between cells to improve clustering and quantification, and to analyze expression at the resolution supported by the data, i.e., as groups of distinguishable isoforms. All of these tools will be released as high-quality, open-source software.