Study Of Drug-target Interaction Network Prediction Methods

The first key step in the current new drug discovery process is to find, identify and prepare the target proteins for drug molecules. In particular, several classes of proteins such as enzymes, ion channels, G-protein-coupled receptors (GPCR) and nuclear receptors represent the vast majority of current drug targets for new drug development. However, due to the impact of throughput, accuracy and cost, it is difficult to elucidate these potential drug-target interactions by the application of traditional experimental methods. Therefore, there is an urgent need to develop effective computational method so as to help researchers finding the regular pattern of drug-target interactions and provide a complementary and supportive evidence for the experimental study. In this dissertation, according to reliable data sources, we developed feature selection, supervised learning, semi-supervised learning and network topology analysis methods to effectively predict potential drug-target pairs by integration of a variety of biological information such as protein homologous similarity, molecular descriptors information, protein function information, chemical structure similarity, functional groups and drug-target network topology information. The main works and contributions for this dissertation are introduced as follows:1. A semi-supervised learning method was introduced to maximize the use of existing labeled data to infer a large amount of unlabeled data for predicting potential drug-target interactions from chemical space and genomic space in four known classes of drug-target network. Results showed that the performance of drug-target interaction predciton was improved by comparing different protein homologous algorithms and chemical structural algorithms.2. An improved bipartite graph learning approach was developed to predict drug-target interaction network for four classes of enzymes, ion channels, G-protein coupled receptors (GPCR) and nuclear receptors. Through performing feature selection method for biological features, this kernel-based approach was used to detect unkown drug-target interactions. Results showed that the performance of the proposed approach was superior to the bipartite graph learning approach. Some predicted drug-target pairs were also confirmed in curated databases.3. According to the known drug-target interaction dataset of13kinds of disease, a graph-based semi-supervised learning method was proposed. This method extracted the drug-target network toplogy information and integrated multiple biological features so as to infer new drug-target pairs from the known drug-target network. Results showed that this method had better performance than the existing biparitie graph local model method (BLM) and semi-supervised learning method (NetlapRLS).