Prediction Of Drug-target Interaction And Drug Response In Cell Lines

Prediction of new targets for drugs and the sensitive or resistive response of drugs are very important,which can not only help finding new functions for drugs,facilitate the understanding of drug action mechanism,find the individual differences in drug efficacy,which in turn can further help to understand the reason for off-target,promote precision medicine.As drug development is an expensive and time-consuming process with extremely low success rate,so the research of drug reposition,also known as finding new uses for old drug and the drug response for personal disease are important for precision medicine.This paper will integrate multiple omics data,such as proteome,genome,pharmacology,disease phenotype and metabolic pathways,to conduct the thorough research for drug-target interaction and drug-cell line response methods,thesis main contributions are as follows:1.For solving the problem that supervised classification-based methods only introduce limited feature of drug and protein,and combine multi-features to predict drug-target interaction,and the prediction accuracy can be further improved,we proposed one novel method for predicting drug-target interaction by using the decision template(namely DT-all).DT-all first introduces Gene Ontology and pathway annotation feature to compute the target similarity features,coding the target protein sequence,and extracting compound structure and Anatomical Therapeutic Chemical classification-based(ATC-based)code for drugs.Then target features and drug features are combined to input into KNN classifiers.Finally,decision template is performed to integrate the output of multiple classifiers,obtaining the final score for each drug-target sample.In 5 fold cross-validation test,the results show that the performance of DT-all outperforms other state-of-the-art supervised classification-based methods.The Gene Ontology and pathway annotation features can further improve the prediction accuracy.2.For solving the problem that network-based methods do not utilize the heterogeneous network information and the prediction accuracy can be further improved,we proposed a new method for predicting drug-target interaction by using the network-based label propagation with mutual interaction information derived from heterogeneous network(namely LPMIHN).Firstly,LPMIHN integrates the drug's chemical similarity,drug topological similarity,target sequence similarity,target topological similarity and the known DTI bipartite network to establish the drug-target heterogeneous network.Then,fuses the target(or drug)label and the known DTI network information to obtain the initial label information of drug(or target).Finally,for a specific drug(or target),run label propagation on drug similarity network(or target similarity network)and target similarity network(or drug similarity network)sequentially to predict the potential targets(or drug).Comparison with other recent state-of-the-art network-based methods and the results show that LPMIHN achieves the best results in terms of AUC and AUPR.In addition,many of the promising drug-target pairs predicted from LPMIHN are also confirmed on the latest publicly available drugtarget databases.3.In view of the limited integration of heterogeneous information,the drug and target similarity networks cannot be constructed effectively,which in turn affects the prediction accuracy.We propose a multiple kernel learning and Bi-random walk-based algorithm for predicting drug-target interaction(namely MKLC-BiRW).Firstly,MKLC-BiRW integrates diverse drug-related and target-related heterogeneous information source with multiple kernel learning for generating the comprehensive drug and target similarity matrices.Secondly,adopts the cohesiveness of the clusters of drug-target interaction network,drug-drug interaction network and protein-protein interaction network to adjust the drug and target similarity matrices,respectively.Thirdly,using the results of statistical significance relationship between pairwise drug/target similarity and the number of common targets/drugs to verify the drug/target similarity matrices,and builting a heterogeneous network by incorporating the drug/target similarity networks with known drug-target interactions.Finally,with the drug-target heterogeneous network,Bi-random walk algorithm is adopted to infer the potential drug-target interactions.Compared with other methods,it is shown that MKLC-BiRW method has reliable performance for predicting the potential drug-target interactions.4.In order to reveal the causes of drug sensitivity or resistance to cell lines from pharmacological and genomic perspectives,and further improve the prediction accuracy,we proposed a low rank triple matrix factorization based method for drugcell line response prediction(namely TMF).TMF first constructs drug feature matrix and cell line feature matrix based on the chemical substructure fingerprint and the gene expression profiles.Then,normalizes the drug response data and constructs a drug-cell line response matrix in which some elements are unknown.Finally,a low rank triple matrix factorization method is designed by connecting drug feature matrix,cell line feature matrix and drug-cell line response matrix together,to predict the unknown drugcell line response.By testing on GDSC dataset,the results show that the performance of TMF outperforms other state-of-the-art matrix factorization-based methods,such as KBMF and SRMF.On the other side,by using the projection matrix obtained from TMF method,we found important drug feature and cell line feature for response prediction,which can help to understand the machanism of sensitivity/resistive response between drugs and cell lines.