Features of Gene Extraction by Nonlinear Support Vector Machines in Gene Expression Analysis

Statistical analysis on gene expression data from DNA microarray has enabled us to extract information from tissue and cell samples. Comparing two classes of gene expression datasets (e.g. datasets from normal tissues and cancerous tissues), we first choose discriminative genes, which have significantly different expression values between two classes and characterize each class. In the most case of statistical filter methods such as t-test, the chosen genes tend to be strongly expressed in one class and weakly in the other class. However, the lower ranked genes by the simple filter methods may also include discriminative and informative genes. The expression values of such genes have complicated distributions. Suppose there are two classes whose distribution is as depicted in Figure 1: there are two peaks of frequency of gene expression values in one class, while the peak in the other class is located between them. In this case, although such a gene is of importance, it is low ranked and thus discarded by conventional filter methods because of its low inner-class average and its high deviation. In supervised classification problems, various wrapper methods such as Recursive Feature Elimination (RFE) [1] are proposed for feature selection. Since the main purpose of the wrapper method is to improve the performance of the classification algorithm, the genes chosen by the method have not been paid attention, especially in the nonlinear classification. In this paper, we use the wrapper methods based on a nonlinear classification algorithm in order to extract the discriminative genes that difficult to be extracted by conventional filter methods. RFE method based on nonlinear Support Vector Machines (SVMs) [2] is employed to this end because it is successfully applied to classification of gene expression data. We investigate the genes extracted by the RFE method based on SVMs with gaussian kernel function to indicate that it can extract discriminative genes which are not chosen by conventional filter methods.