MIB: Using mutual information for biclustering gene expression data

Result of any biclustering or clustering algorithm depends on the choice of the similarity measure. Most of the biclustering algorithms are based on Euclidean distance or correlation coefficient. These measures capture only linear relationships between the genes but nonlinear dependencies may exist amongst them. In this paper we propose an approach using mutual information for biclustering gene expression data. Mutual information is a more general measure to investigate relationships (positive, negative correlation and nonlinear relationships as well). To the best of our knowledge, none of the existing algorithms for biclustering have used mutual information as a similarity measure between two genes. We obtained biclusters from the gene expression data of Arabidopsis thaliana and compared our biclusters with those obtained by two other algorithms namely ISA and BIMAX. Biological significance of the biclusters was checked using GO database. It was found that the genes belonging to our biclusters were significantly enriched with GO terms with better p values as compared to the genes of the biclusters obtained by the other two algorithms. To further investigate the biclusters, we studied the promoter regions of the genes belonging to a bicluster for common patterns/transcription factor binding sites (TFBS) or motifs. Promoter regions of the genes of most of our biclusters were found to have a common motif patterns which existed in the motif database of Arabidopsis thaliana. Also, the motifs extracted from our biclusters had better E values than those of others. Thus reconfirming that use of mutual information as a similarity measure will produce better biclusters.