Statistical identification of biclusters in gene expression data

A bicluster of a gene expression dataset is a subset of genes which exhibit similar expression patterns along a subset of conditions. Given a gene expression matrix, we search for submatrices that are tightly coregulated according to some scoring criterion. We do not require the identified submatrices to be disjoint or to cover the entire matrix; instead we wish to build a diverse collection of submatrices that will capture all the significant signals in gene expression data. We believe that the size of the bicluster should be small compared to the size of the gene expression data matrix. So our approach finds biclusters by starting from small tightly co-regulated submatrices and adding more rows and columns to them. Our algorithm has three steps. First, we generate a set of high quality bicluster seeds based on a partition based clustering technique. In the second phase, these bicluster seeds are enlarged by adding more genes and conditions. In the third phase, we find the p-values of the biclusters produced for statistical validation.