An ensemble greedy algorithm for feature selection in cancer genomics

Signature learning from gene expression consists in selecting a subset of molecular markers which best correlate with prognosis. It can be cast as an optimization-based feature selection problem. Here we use as optimality criterion the separation between survival curves of clusters induced by the selected features. We address some important problems in this fields such as developing an unbiased search procedure and significance analysis of a set of generated signatures. We apply the proposed procedure to the selection of gene signatures for Colon Cancer prognosis by using a real data-set.