Modelling p-value distributions to improve theme-driven survival analysis of cancer transcriptome datasets