Challenges and strategies in transcriptome assembly and differential gene expression quantification. A comprehensive in silico assessment of RNA-seq experiments

Transcriptome Shotgun Sequencing (RNA-seq) has been readily embraced by geneticists and molecular ecologists alike. As with all high-throughput technologies, it is critical to understand which analytic strategies are best suited and which parameters may bias the interpretation of the data. Here we use a comprehensive simulation approach to explore how various features of the transcriptome (complexity, degree of polymorphism π, alternative splicing), technological processing (sequencing error ε, library normalization) and bioinformatic workflow (de novo vs. mapping assembly, reference genome quality) impact transcriptome quality and inference of differential gene expression (DE). We find that transcriptome assembly and gene expression profiling (EdgeR vs. BaySeq software) works well even in the absence of a reference genome and is robust across a broad range of parameters. We advise against library normalization and in most situations advocate mapping assemblies to an annotated genome of a divergent sister clade, which generally outperformed de novo assembly (Trans-Abyss, Trinity, Soapdenovo-Trans). Transcriptome complexity (size, paralogs, alternative splicing isoforms) negatively affected the assembly and DE profiling, whereas the effects of sequencing error and polymorphism were almost negligible. Finally, we highlight the challenge of gene name assignment for de novo assemblies, the importance of mapping strategies and raise awareness of challenges associated with the quality of reference genomes. Overall, our results have significant practical and methodological implications and can provide guidance in the design and analysis of RNA-seq experiments, particularly for organisms where genomic background information is lacking.