Reader : Fabio Vandin

In studying the correlations between genetic mutations and cancer development, the relationship between mutations and patient survival is one of the most significant from both academic and clinical perspectives. Survival analysis being an established field, assessing the correlation between any given mutation or collection of mutations and survival is a straightforward process. This paper seeks to develop techniques to use survival analysis to identify de novo collections of mutations correlated with low survival. This requires exploring a very large search space of collections of genes, which cannot be effectively enumerated with computational resources. Instead local search in the form of steepest ascent hill climbing and a more generalized branching algorithm will be used to iteratively construct strongly correlated sets. Theoretical analysis of the runtimes demonstrate the feasibility of these algorithms under some limited assumptions. Empirical results demonstrate the ability of these algorithms to recover sets strongly correlated with survival. Some evidence is presented indicating that these techniques, while somewhat effective at present, will be significantly more so when cancer genomic datasets reach two to three times there current size.