The paired end, stranded reads were quality filtered with Trimgalore to avoid reads with low phred score to rule out false alignments. Reads were mapped to hg19 and Gencode v.18 as a gene model using Tophat2 with default options and first strand library type3637. The mapped reads were counted with HTSeq to known genes from Gencode v.18 using the union model coupled with reversed strand (http://biorxiv.org/content/biorxiv/early/2014/08/19/002824.full.pdf). The gene counts were normalized to counts per million (cpm) reads and to the upper quartile using the edgeR bioconductor package38. This package measures the dispersion towards a trended mean and has been shown to be sensitive to outliers. Genes with less than one cpm in at least three samples were eliminated from further analysis recommended from previous methods39. Differential expression was measured with edgeR and the P-values were adjusted with Benjamini-Hochberg multiple testing. Genes were deemed significant if they had a FDR less than 5%.
