Modeling biclustering as an optimization problem using mutual information

Most of the biclustering algorithms for the analysis of high dimensional gene expression data use some distance measure or correlation coefficient between a pair of genes as the similarity measure. These measures capture only linear relationships between the genes but non linear relationships may exist amongst them. Mutual information is a more general measure to investigate relationships (positive, negative correlation and non linear relationships as well). Biclustering problem has been modeled as an optimization problem and a polynomial time solution has been proposed for it. In this paper we use Mutual Information to model Biclustering as an optimization problem. Results on gene expression data of Arabidopsis Thaliana show that our method produces different yet biologically significant biclusters. It was found that our biclusters had better p values as compared to the biclusters obtained by the other existing algorithms. Also promoter region of the genes of most of the biclusters were found to have common motif patterns.