Drug Repositioning Research On Multi-omics Heterogeneous Biological Networks

The development of new drugs has the problems of high investment,high risk,time-consuming and laborious.It usually takes 10-15 years of research and 0.8 billion dollars from discovery to approval of a new drug.Besides,a drug needs to go through molecular exploration,animal experiment,clinical trials research and so on,while the success rate is extremely low.Drug repositioning,also known as old drugs for new use,refers to the discovery of new indications or new uses of marketed drugs,which can significantly shorten the drug development cycle,reduce pharmaceutical costs and avoid risks,which has gradually become an important strategy of drug development.The main strategies of computational analysis of drug repositioning can be summarized as drug-target association prediction,drug-disease association prediction,and drug-drug interaction prediction.In this thesis,we mainly focus on designing models on the strategy of drug-target association analysis and drug-disease association analysis.We use the link prediction algorithm in heterogeneous network to realize drug repositioning.Specially,this thesis does the following three aspects of work and innovation:(1)In recent years,with the rapid development of bioinformatics,systems biology,deep learning and big data,the accumulation of biological data such as massive drugs,diseases and targets have provided rich biological features.In this thesis,we integrate 15 types of chemical,genomic,phenotypic,and cellular networks to build a multiomics heterogeneous biological network.In order to learn deeper understanding of different networks,we use the multilayer deep autoencoder to obtain low dimensional vector representation for each node in the network.Considering the imbalance of positive and negative samples in biological data,we introduce PU matrix completion to find the best projection from drug space onto target space.(2)When learning the representation of nodes in the network,most research only consider the proximity within the second order between neighbors.Prior works have indicated that,besides the pairwise edges,the high-order proximity between nodes is of vast importance in capturing the underlying topological structure of the network.In this thesis,we learn low dimensional vector representation of nodes in 15 networks using an Arbitrary-order Proximity Embedded Deep Forest approach,which can preserve arbitrary-order proximity features for nodes.Then we utilize a tree model with high classification accuracy to complete link prediction.Specifically,we treat the link prediction task of drug-target as binary classification problem in machine learning and use the deep forest model for classification.(3)In the first work,we use the multilayer deep autoencoder to extract features from each network separately,however,we have not considered the association information between different networks,which may lead to information lose when integrating the features of all networks.Learning from the idea of multi-model learning,we improve the model through the use of multi-model deep autoencoder,for the feature extraction and fusion of different drug-related networks,thus we can learn a highquality feature representation relevant to all networks.Then we consider the drugdisease link prediction problem as recommendation problem,and recommend drugs with high prediction scores for different diseases via a collective variational autoencoder.