The Research Of Drug-target Interaction Prediction And Its Application On Drug Repositioning

Drug research and development needs a large amount of funds and time,and is high risk and low-rate success.According to statistics,developing a new drug will take 10-15 years and 0.8-1.5 billion dollars from determining the idea to enter the market.Nobel Lau-reate James Black has said that the most solid foundation for the research and development of new drugs is starting from old drugs.Therefore,more and more pharmaceutical compa-nies intend to mine new clues of existing drugs by filtering the existing drug molecules to accelerate drug development.Determining whether a drug interacts with a target is one of the key stages of accel-erating drug research and development.Therefore,this thesis develops different models according to different tasks to predict possible drug-target interactions,and thus reposition existing drugs and targets and infer new clues of treatment for related diseases.Firstly,this thesis develops a negative screening method,NDTISE(Negative Drug-Target Interaction Sample Extraction),to extract high-quality negative drug-target inter-action samples.Firstly,this thesis fuses various biological information related to drugs and targets and represents a drug-target interaction as a vector,and extracts feature subsets based on discriminative ability score of each feature in positive sample set and unlabeled sample set;Then,it extracts reliable negative samples based on PU learning and multi-classifier combinational idea,and computes representative positive and negative prototypes and sim-ilarity weights of remaining ambiguous samples;Finally,it builds an SVM-based classi-fication model to classify unlabeled drug-target pairs.To evaluate the performance of thr proposed model,it conducts a series of comparative experiments,and focus on the per-formance's of three negative sample screening methods on 6 classical classifiers:random selection method,NCPIS,and NDTISE.The results indicates that the proposed method can effectively screen high-quality negative drug-target interactions.Then,this thesis exploits a drug-target interaction prediction method combining Neigh-bor Interaction profile inferring,Non-negative matrix factorization,Discriminative low-rank representation,and Sparse representation classification(PreNNDS),to mine asso-ciation information for new drugs and targets.The related researches have shown that known neighbor information plays an important role in drug-target interaction identification.Therefore,this thesis presents a drug-target interaction prediction framework,PreNNDS,considering the sparse,low-rank,and nonnegative properties of drug-target interaction ma-trix.PreNNDS integrates known Neighbor Interaction Profiles(NIPs),Nonnegative Matrix Factorization(NMF),Discriminative Low-Rank Representation(DLRR),and Sparse Repre-sentation Classification(SRC)into a unified framework.PreNNDS first computes associated probability for each drug-target pairs based on NIPs,and then extracts feature matrices of drugs and target based on NMF method,the following builds DLRR-based optimization model,finally predicts interaction information for new drugs or targets based on SRC.The experimental results indicate that PreNNDS can effectively infer the correlation information for new drugs or targets.Following,the thesis proposes a Multi-Information fusion method combining Norm idea(NormMulInf)based on various biological properties of drugs and targets.Firstly,NormMulInf computes drug similarity matrix and target similarity matrix based on chemi-cal structure similarities of drugs,sequence similarity of target proteins,and drug topologi-cal similarities and target topological similarities in drug-target interaction network.Then,it utilizes a small number of known labeled data and a large number of unlabeled data and presents two prediction models,NormDrug(Model based on Drug similarity and Norm idea)and NormTarget(Model based on Target similarity and Norm idea),based on robust PCA model;And it applys exact ALM algorithm and solves the models based on minimiz-ing nuclear norm and l1 norm;Finally,NormMulInf is presented based on the NormDrug and NormTarget to complete the missing data.The results from a series of comparative experiments shows the performance of NormMulInf.Finally,after determining the performances of the above three methods,the thesis repositions the existing drugs and targets based on negative samples extracted by NDTISE,predicts the correlation information for new drugs and targets,infers new treatment clues of Alzheimer's disease,and gives 100 drug-target interactions with the highest scores.In general,this thesis combine the related theories and algorithms on machine learning and exploit different drug-target interaction prediction models from negative sample selec-tion,prediction based on the simple source data,and multi-information fusion according to different tasks under different conditions.After determining the performance of the pro-posed model,the thesis repositions existing drugs and targets.Further retrieving predictive results on public databases and relevant literatures indicates that these prediction results is worth of further biomedical experiment validation.