Robust Weighted Kernel Logistic Regression to predict gene-gene regulatory association

Gene-gene associations are usually inferred from correlations between pairs of genes in different types of biological data such as microarray expression measurements. However coexpression (or correlation) based associations are meaningful only when data sets share the same experimental conditions. In addition, correlation does not indicate a regulatory relationship between two genes. In this work, we adopt another approach to identify connected (with regulatory relationship) pairs of genes utilizing genome-wide expression data with a wide range of different experimental conditions. The number of connected pairs of genes is typically a very small number compared to the total number of pairs in the whole genome. Thus, we can assume that gene-gene connection is a rare event that can be predicted using special classification algorithms. The algorithm used here is called Rare Event-Weighted Kernel Logistic Regression (RE-WKLR). The features that define each pair of genes are the moments of joint probability distribution of expression levels of these two genes. This approach is applied to Saccharomyces cerevisiae genome. The accuracy of RE-WKLR in predicting gene-gene connections is compared with that of Support Vector Machines (SVM) and is found to be higher than that of SVM.