A novel multiclass gene selection method based on SVM/MLP cross validation

gene selection is one of the major challenges of biochip technology for resolution of curse of dimentionality which occurs especially in DNA microarray dataset where there are more than thousands of genes and only a few experiments (samples), and for gene diagnosis where only a gene subset is enough for diagnosis of diseases. This paper presents a gene selection method by training linear SVM (support vector machine)/nonlinear MLP (multi-layer perceptron) classifiers and testing them with cross validation for finding gene subset which is optimal/suboptimal for diagnosis of binary/multiple disease classes. The process is to select genes with linear SVM classifier incrementally for the diagnosis of each binary disease class pair, by testing its generalization ability with leave-one-out cross validation; the union of them is used as initialized gene subset for the discrimination of all the disease classes, from which genes are deleted one by (one decrementally by removing the gene which brings the greatest decrease of the generalization power after the removal, where generalization is measured by leave-one-out and leave-4-out cross validation. For real DNA microarray data with 2308 genes and only 64 labelled samples belonging to 4 disease classes, only 6 genes are selected to be diagnostic genes. The diagnostic genes are tested with 6-2-4 MLP with both leave-one-out and leave-4-out cross validation, resulting in no misclassification.