The problem of identifying interacting genes that jointly are associated with a phenotype is considered. When the number of features are extremely large compared to the number of samples, there may be several subsets of features that provide acceptable levels of predictability. This is particularly true in cancer genomics, where we are interested in finding functionally related gene sets likely to jointly drive cancer phenotypes. In this paper, a novel game theoretic solution is proposed by modeling genes as players of a Coalition Game. This method discovers and develops informative gene subnetworks by integrating gene expression profiling of cancer tissues with protein-protein interaction (PPI) networks. These subnetworks are gradually developed by selective addition of candidate genes that present maximal Shapely values in coalition with subnetworks of genes. We applied the proposed algorithm to an ovarian cancer dataset (N = 201), in order to identify optimal subnetworks that can predict cancer progression risk in response to platinum-based therapy. We show improved predictive power of the proposed method when compared to state-of-the-art feature selection methods, with the added advantage of identifying potentially functional gene subnetworks that may provide insights into the mechanisms underlying cancer progression.