Biomolecule Network Analysis And Its Application In Disease-drug Prediction

Due to the development of high-throughput biotechnology,the amount of data of biological systems is growing rapidly,the demand for mining effective biological information from massive data becomes increasingly urgent.In general,a disease is caused by the interactions between molecules in complex biological systems,which motivates us to explore the potential factors causing the problem from biological data.The network-based approaches can boost the study of complex diseases and solve biomedical problems,e.g.,clinical diagnosis and new utilization of old drugs,both conveniently and efficiently.In this thesis,single homogeneous and complex heterogeneous biomolecular networks are studied by using the interactions between molecules in biological complex systems.Complex network control theory and network embedding methods are proposed to address problems of key gene identification of a disease and the prediction of disease-drug relationships.We provide a theoretical basis for biomedical technology.The main work is as follows:(1)Predicting the critical disease genes based on the protein-protein interaction network.For a single homogeneous protein interaction network,capturing the state of the entire network by using complex network control theory allows us to find potential disease-causing genes from massive protein interaction data.To explore the key genes,we use five cancer-related protein-protein interaction networks,which contain 15,474 proteins and 170,631 edges in total.Then we utilize the Minimum Dominating Set(MDS)of maximum connected subgraph of network to select the genes that always belong to the MDS as the candidate key genes.By using the tumor-related pathways and essential tumor gene sets,we find that the candidate key genes are clustered in these gene sets,which indicates the effectiveness of the methods based on MDS.In addition,we build a comprehensive centrality method,and rank the candidate genes thorough this method.We select the top ranked genes as candidate biomarkers.Furthermore,we evaluate the probability of top ranked genes being biomarkers according to the network structural analysis and the enrichment of the somatic mutation.In summary,this study may shed light on the application of complex network control theory in biomedicine.(2)Research on disease-drug prediction based on biomolecule heterogeneous network.We propose different embedding algorithms on the complex heterogeneous disease-gene-drug interaction network by considering different ways of random walk sampling methods,including the commonly used embedding algorithms,such as Deepwalk and Node2 vec.By fully training the interaction of disease-gene,drug-gene and gene-gene,we learn the potential disease-drug connection and potential indications.This method solves the problem of heterogeneous link prediction on heterogeneous networks.Meanwhile,the network embedding method is more efficient than the previous method based on the shortest path,and greatly improves the efficiency of disease drug prediction based on biological data.Moreover,it can effectively shorten the cost of time of looking for disease-drug potential relationships.