Additive noise analysis on microarray data via SVM classification

Abstract

Microarray technology has been broadly used for monitoring the expression levels of thousands of genes simultaneously, providing the opportunities of identifying disease-related genes by finding differentially expressed genes in different conditions. However, a great challenge of analyzing microarray data is the significant noise brought by different experimental settings, laboratory procedures, genetic heterogeneity among samples, and environmental variations among different patients, and so on. This paper attempts to analyze the influence of these noises on each gene by measuring the changes of classification performance. We assume each gene in microarray data includes an independently distributed unknown uniform noise. Thus, we add a compensated noise back to each gene and test whether the classification accuracy of a linear support vector machine (SVM) improves. If the accuracy does increase, then we believe such noise does exist and degenerate the relation of this gene to the disease status. Through extensive experiments on several public microarray data, we found such added noises can improve the classification accuracy in several genes and the results are relatively consistent, indicating our method can be used to analyze the noise pattern in microarray experiments, and also discover potential important gene markers.