Comparing Segmentation Methods for Genome Annotation Based on RNA-Seq Data

Transcriptome sequencing (RNA-Seq) yields massive data sets, containing a wealth of information on the expression of a genome. While numerous methods have been developed for the analysis of differential gene expression, little has been attempted for the localization of transcribed regions, that is, segments of DNA that are transcribed and processed to result in a mature messenger RNA. Our understanding of genomes, mostly annotated from biological experiments or computational gene prediction methods, could benefit greatly from re-annotation using the high precision of RNA-Seq. We consider five classes of genome segmentation methods to delineate transcribed regions, including intron/exon boundaries, based on RNA-Seq data. The methods provide different functionality and include both exact and heuristic approaches, using diverse models, such as hidden Markov or Bayesian models, and diverse algorithms, such as dynamic programming or the forward-backward algorithm. We evaluate the methods in a simulation study where RNA-Seq read counts are generated from parametric models as well as by resampling of actual yeast RNA-Seq data. The methods are compared in terms of criteria that include global and local fit to a reference segmentation, Receiver Operator Characteristic (ROC) curves, and coverage of credibility intervals based on posterior change-point distributions. All compared algorithms are implemented in packages available on the Comprehensive R Archive Network (CRAN, http://cran.r-project.org). The data set used in the simulation study is publicly available from the Sequence Read Archive (SRA, http://www.ncbi.nlm.nih.gov/sra). While the different methods each have pros and cons, our results suggest that the EBS Bayesian approach of Rigaill, Lebarbier, and Robin (2012) performs well in a re-annotation context, as illustrated in the simulation study and in the application to actual yeast RNA-Seq data. This article has supplementary material online.