RNA-Sequencing (RNA-seq) has enabled the in-depth study of the transcriptome, becoming the primary research method in the field of molecular biology. The typical aim of RNA-seq is to quantify and detect differentially expressed (DE) and differentially spliced (DS) genes. Numerous methodologies and tools have been developed in recent years to assist in analyzing RNA-seq data. However, it is difficult for researchers to decide which methods or strategies they should adopt to optimize the analysis of their datasets.

In this Thesis, in Study I, we applied the gene-level DE analysis approach to detect the androgen-regulated genes between cancerous and benign samples in 48 primary prostate cancer patients. Combined with other measurements from the same samples, our analysis indicated that patients having TMPRSS-ERG gene fusion had distinct intratumoral androgen profiles compared to TMPRSS-ERG negative tumors. However, the DE can remain undetected when the expression varies across the gene due to reasons such as alternative splicing. Hence, to account for this problem, an alternate analysis approach has been suggested in which the statistical testing of lower feature levels (e.g. transcripts, transcript compatibility counts, or exons) is performed initially, followed by aggregating the results to the gene level. In Study II, we tested this alternate approach on these lower features and compared the results to those from the conventional gene-level approach. In the alternate approach, two methods (Lancaster method and empirical brown method (ebm)) were tested for aggregating the feature-level results to gene-level results. Our results suggest that the exon-level estimates improve the detection of the DE genes when the ebm method is used for aggregating the results. Accordingly, R/Bioconductor package EBSEA was developed using the winning approach.

RNA-seq data can also be used to find DS events between conditions. However, the detection of DS is more challenging than the detection of DE. In Study III, a comprehensive comparison of ten DS tools was performed. We concluded that exonbased and event-based methods (rMATS and MAJIQ) performed overall best across the different evaluation metrics considered. Furthermore, we observed overall low concordance between the results reported by the different tools, making it recommendable to use more than one tool when performing DS analysis, and to concentrate on the overlapping results.