A hybrid knowledge based-clustering multi-class svm approach for genes expression analysis

This study utilizes Support Vector Machines (SVM) for multi-class classification of a real data set with more than two classes. The data is a set of E. coli whole-genome gene expression profiles. The problem is how to classify these genes based on their behavior in response to changing pH of the growth medium and mutation of the acid tolerance response gene regulator GadX. In order to apply these techniques, first we have to label the genes. The labels indicate the response of genes to the experimental variables: 1-unchanged, 2-decreased expression level and 3-increased expression level. To label the genes, an unsupervised K-Means clustering technique is applied in a multi-level scheme. Multi-level K-Means clustering is itself an improvement over standard K-Means applications. SVM is used here in two ways. First, labels resulting from multi-level K-Means clustering are confirmed by SVM. To judge the performance of SVM, two other methods, K-nearest neighbor (KNN) and Linear Discriminant Analysis (LDA) are implemented. The Implementation of Multi-class SVM used one-against-one method and one-against-all method. The results show that SVM outperforms KNN and LDA. The advantage of SVM includes the generalization error and the computing time. Second, different from the first application, SVM is used to label the genes after it is trained by a set of training data obtained from K-Means clustering. This alternative SVM strategy offers an improvement over standard SVM applications.