Network Kernel SVM for microarray classification and gene sets selection

The importance of network-based approach to identifying biological markers has been increasingly recognized. Lots of papers indicated that genes in a network tend to function together in biological processes, so taking full advantage of the biological observation can improve the performance of microarray classification. However, lots of SVM methods don't consider this situation during their classifier building. The main idea of this paper intends to embed the information of gene networks into a new SVM learning framework. Based on a new regularization, we propose a novel method, Network Kernel SVM (NK-SVM), for binary classification problem and gene sets selection. By constructing some special kernel matrixes from the prior information of gene network, the new NK-SVM method makes the genes in the same set to be selected (or eliminated) together. The numerical experiments on a real microarray application show that the proposed method tends to provide a better performance than other methods on gene sets selection.