Introduction

The construction of transcriptional regulatory networks is of utmost importance to decipher the regulation of biological processes. Human transcriptional regulatory networks are known to be cell-type specific. Recently, some computational methods (including Centipede and PIQ) have been proposed in the literature to predict cell-type specific transcriptional factor binding sites (TFBS's) from DNase-seq data and global motif information. Regulatory networks can be constructed by correlating the TFBS's with the genes in their neighborhoods. However, a TFBS does not always imply regulatory control in the transcriptional process. In this project, we attempt to refine the regulatory networks by integrating gene expression data with the predicted TFBS's. More specifically, similariy between cell types is estimated by using the expression data from ENCODE and a Markov random field (MRF) model is defined to connect the corresponding regulatory relationships of different cell types based on the similarity. The MRF is then optimized to obtain refined regulatory networks for 110 ENCODE cell lines.