The combined use of gene expression profiles and protein-protein interaction networks has shown remarkable successes in the prediction of breast cancer metastases. Nevertheless, as a primary step of network-based methods, the problem of effectively identifying predictive subnetwork markers remains a great challenge. Typically, existing methods use greedy search algorithms to search for subnetworks. This strategy, though efficient in time complexity, may fail in finding the optimal subnetwork markers and accordingly impair the performance of the successive learning machines. In this paper, we propose a genetic algorithm to improve the subnetwork markers that have been identified by an existing greedy search method. We demonstrate that the discriminative power of the optimized subnetwork markers are significantly higher than the original subnetwork markers, and we show that higher classification performance can be achieved when using the optimized subnetworks as predictive features via six popular machine learning approaches (logistic regression, support vector machine, decision tree, Adaboost, random forest and Logitboost). According to the comparison between different classification approaches, Logitboost with the optimized subnetwork markers shows the highest classification performance and optimal reproducibility for identifying breast cancer metastases.