An Information Theoretic Algorithm for Mining and Ranking Phenotype-specific Sub-networks from Multi-class Gene Expression Data

Abstract

There have been extensive studies for inferring transcriptional network from omics data. However, how to utilize networks for specific research projects has not been well established. One of the main hurdles is lack of algorithms for mining biological sub-networks. Existing graph mining algorithms do not consider features of the transcriptional network and they are not effective to obtain biologically meaningful results. In this paper, we define the biological sub-network mining problem and present a new graph mining algorithm that mines and ranks phenotype specific sub-networks of transcriptional regulatory networks constructed from multi-class gene expression data. Our contributions in this paper on the computational side are two folds. First, we suggest a complete research paradigm of utilizing omics data to construct networks and then elucidate s ub-networks that distinguish phenotypes or disease states. Second, we developed an information theoretic algorithm for mining phenotype specific sub-networks. Our contribution on the bio/medical side is that our TF-module based analysis determined biological pathways (cell cycle: M-phase, cell adhesion molecules) related to the phenotype (breast tumor grade) by identifying activation/suppression of specific target genes (TGs) by the combination of multiple transcription factors (TFs). Expression levels of TGs clearly shows correlation between activation/suppression of these pathways and tumor grades. When we used all genes, pathway activation or suppression was not obvious, which shows the effectiveness of our algorithm. Our TF-centric pathway activation/suppression analysis technique is applicable to and useful for many other studies.