Linear Coherent Bi-cluster Discovery via Line Detection and Sample Majority Voting

Discovering groups of genes that share common expression profiles is an important problem in DNA microarray analysis. Unfortunately, standard bi-clustering algorithms often fail to retrieve common expression groups because (1) genes only exhibit similar behaviors over a subset of conditions, and (2) genes may participate in more than one functional process and therefore belong to multiple groups. Many algorithms have been proposed to address these problems in the past decade; however, in addition to the above challenges most such algorithms are unable to discover linear coherent bi-clusters--a strict generalization of additive and multiplicative bi-clustering models. In this paper, we propose a novel bi-clustering algorithm that discovers linear coherent bi-clusters, based on first detecting linear correlations between pairs of gene expression profiles, then identifying groups by sample majority voting. Our experimental results on both synthetic and two real datasets, Saccharomyces cerevisiae and Arabidopsis thaliana , show significant performance improvements over previous methods. One intriguing aspect of our approach is that it can easily be extended to identify bi-clusters of more complex gene-gene correlations.