Statistical models for biological and technical variation in RNA sequencing

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): Since the invention of microarrays, measuring genome-wide gene expression is one of the most common experiments performed by molecular biologists. Gene expression analysis is also widely used in clinical applications to discover the molecular architecture of disease or to develop prognostic and predictive signatures. RNA-sequencing (RNA-seq) has become the preferred technology for making expression measurements due to declining costs and because RNA-seq is flexible enough to measure expression in regions not previously annotated as genes and to measure the abundances of multiple transcripts for individual genes. Now that RNA-seq data can be collected inexpensively and processed in experiments with replicates, a major challenge is statistical modeling and interpretation of results from RNA-seq experiments. Our proposal will tackle three key practical challenges in RNA-seq data analysis: (1) estimation and removal of hidden artifacts, (2) statistical models for differential expression scanning that d not rely on annotation or assembly, and (3) robust statistical models to correct ambiguous, variable, and unidentifiable assemblies, with specific application to the most popular computational RNA-seq software, Cufflinks. The first aim extends our batch discovery and removal methods to RNA-sequencing data by modeling within gene and spatial dependence in expression estimates that lead to heavily biased artifact estimates and reduced power. The second aim develops a statistical framework for first identifying regions of differential expressio at base-pair resolution, then associating these regions with known genomic landmarks or annotation as a lightweight and accurate scanning approach. This approach builds on the most mature statistical methods for RNA-seq analysis but does not rely on annotation to define transcriptional units such as genes or exons, allowing for unbiased discovery of differential expression. The third aim develops a statistical normalization and analysis framework that addresses the most egregious artifacts and limitations of the inherently ambiguous transcript assembly process. We will work closely with the developers of the most popular RNA-seq assembly software, Cufflinks to integrate our developments into that software suite. By modeling variation across genes using functional regression and in the transcript assembly process using hierarchical models we will reduce the number of false positives and increase the reproducibility of alternative transcript differential expression results. The statistical methods we develop will e packaged in freely available open source software that is designed to interact with downstream Bioconductor packages for summarization and visualization such as IRanges or Genominator. The result of this proposal will be a modular, integrated pipeline for analyzing RNA-seq data from raw reads produced by the sequencing machine to easily summarized and visualized tables of robust, interpretable, and reproducible results - thereby increasing the number and range of applications of RNA-seq in molecular biology and medicine.