Mining Low-Variance Biclusters to Discover Coregulation Modules in Sequencing Datasets

High-throughput sequencing (CHIP-Seq) data exhibit binding events with possible binding locations and their strengths, followed by interpretation of the locations of peaks. Recent methods tend to summarize all CHIP-Seq peaks detected within a limited up and down region of each gene into one real-valued score in order to quantify the probability of regulation in a region. Applying subspace clustering techniques on these scores can help discover important knowledge such as the potential co-regulation or co-factor mechanisms. The ideal biclusters generated would contain subsets of genes and transcription factors (TF) such that the cell-values in biclusters are distributed around a mean value with very low variance. Such biclusters would indicate TF sets regulating gene sets with very similar probability values. However, most existing biclustering algorithms neither enforce low variance as the desired property of a bicluster, nor use variance as a guiding metric while searching for the desirable biclusters. In this paper we present an algorithm that searches a space of all overlapping biclusters organized in a lattice, and uses an upper bound on variance values of biclusters as the guiding metric. We show the algorithm to be an efficient and effective method for discovering the possibly overlapping biclusters under pre-defined variance bounds. We present in this paper our algorithm, its results with synthetic, CHIP-Seq and motif datasets, and compare them with the results obtained by other algorithms to demonstrate the power and effectiveness of our algorithm.