Abstract
Recently epigenetic study has shown that DNA methylation and miRNAs are highly relevant to the regulation of gene expressions and transcription. Multiple genomics data such as mRNA and miRNA expression and DNA methylation provide comprehensive views of the molecular changes in a biological system under a particular condition. With availability of large protein-protein interaction networks, expression data, and miRNA target databases to identify regulation networks that have significance changes in expression is a challenge problem. In this paper, we present a novel scoring method to perform integration of experimental data systematically, and develop a searching method for inferring maximum-scoring sub-networks using an adaptive Tabu search method based on the characteristics of biological network topology. In the experiments, we tested the methods by applying them to the human amniotic membrane mesenchymal stem cells and its induced neural cell from Linkou Chang Gung Hospital. With mRNA, miRNA and DNA methylation Data, we found that apoptosis-associated factors and STAT related functional pathways participate in the differentiation process of neural cells cross public well-known miRNA target databases (Tarbase, TargetScan, miRanda, PITA and Diana microT). We validated our results related to the neural processes with p-value lower than 0.05 with functional enrichment toolkit and found more significant genes involved in our sub-networks than only differential expression changes between experimental and control data into consideration. We also show that integrating all of the three expression data can help us extract significant network which is strongly associated with neural–related processes. Our approach can extract functional higher-scoring sub-networks and outperforms than those extracted by randomized and searched from randomized seeds in terms of network scores. In our result, it denotes that the predicted targets of miRNA in TargetScan are more reliable than those in PITA database in the sense that it found more neural-process related processes. In the computational side, our method has a better chance to escape the local optima from a greedy algorithm that leads to higher-scoring networks. We also add the biological characteristic of network topology in Tabu search, it help us to extract more biological significant sub-networks. With insufficient number of the genomics data supported, our approach still can discover the significant regulatory networks using prior biological knowledge integration. Our approach successfully integrates genetic and epigenetic expression and protein interaction networks to identify the regulatory networks in a coherent and efficiency manner.