Background: Theme-driven cancer survival studies address whether the expression signature of genes related to a biological process can predict patient survival time. Although this should ideally be achieved by testing two separate null hypotheses, current methods treat both hypotheses as one. The first test should assess whether a geneset, independent of its composition, is associated with prognosis (frequently done with a survival test). The second test then verifies whether the theme of the geneset is relevant (usually done with an empirical test that compares the geneset of interest with random genesets). Current methods do not test this second null hypothesis because it has been assumed that the distribution of p-values for random genesets (when tested against the first null hypothesis) is uniform. Here we demonstrate that such an assumption is generally incorrect and consequently, such methods may erroneously associate the biology of a particular geneset with cancer prognosis.
Results: To assess the impact of non-uniform distributions for random genesets in such studies, an automated theme-driven method was developed. This method empirically approximates the p-value distribution of sets of unrelated genes based on a permutation approach, and tests whether predefined sets of biologically-related genes are associated with survival. The results from a comparison with a published theme-driven approach revealed non-uniform distributions, suggesting a significant problem exists with false positive rates in the original study. When applied to two public cancer datasets our technique revealed novel ontological categories with prognostic power, including significant correlations between "fatty acid metabolism" with overall survival in breast cancer, as well as "receptor mediated endocytosis", "brain development", "apical plasma membrane" and "MAPK signaling pathway" with overall survival in lung cancer.
Conclusions: Current methods of theme-driven survival studies assume uniformity of p-values for random genesets, which can lead to false conclusions. Our approach provides a method to correct for this pitfall, and provides a novel route to identifying higher-level biological themes and pathways with prognostic power in clinical microarray datasets.