Here we show how hallmarks address the problem of gene set redundancy, by illustrating the difference between GSEA performed with the hallmarks and one that uses thousands of gene sets from the MSigDB collections C1-C6. The dataset contains expression data for 200 glioblastoma multiforme (GBM) and two normal brain samples from the Cancer Genome Atlas Research Network (TCGA) (Verhaak et al., 2010). GBM is the most common, most aggressive malignant primary brain tumor in adults (Ostrom et al., 2013). Necrosis, resulting from a limited supply of oxygen and nutrients, is a critical diagnostic feature of GBM (Karsy et al., 2012). We performed standard GSEA on this dataset with the MSigDB v4.0 collections C1-C6 using the samples’ clinical annotation of percentage of necrosis as a continuous phenotype, the Pearson correlation as the ranking metric, and 1,000 permutations of sample labels to estimate significance.
