The Research On Drug Repositioning Prediction Technology Based On Graph Neural Networks And Representation Learning

Research and development of novel drugs cost too much time and money with high risk and low success rate of new drug development.Drug repositioning is to predict different ther-apeutic indications for approved drugs by various technologies,which would help to reduce time,costs and risks of drug development.With gradually accumulation and availability of omics and pharmacoinformatics data,drug repositioning has entered the stage of combining ra-tional design and experimental screening.The computational method on drug repositioning has become an important topic in the field of bioinformatics.Graph neural networks(GNNs)have gradually become one of hot research direction in machine learning.Since molecules can be naturally represented in the form of graph,applying GNNs to drug repositioning to extract the molecular features,and to effectively enhance the learning ability of the model,which improves the prediction accuracy of various tasks on drug repositioning,thereby providing the potential clues for large-scale experimental screening and further reducing costs and research time.This thesis is oriented to drug repositioning prediction technology based on graph neural networks and representation learning.Four kinds of prediction tasks involved in drug reposi-tioning are studied,and the main contributions are as follows:(1)Research on molecular property prediction task.Most of the traditional molecular representations are based on hand-crafted features and heavily rely on biological experimenta-tions,which are often costly and time-consuming.To address the above limitations,we present an end-to-end deep learning framework based on embedding method,named Smi2Vec-Bi GRU,that extracts the characteristics of SMILES sequence by embedding technique,and it adopts a bidirectional gated recurrent unit(Bi GRU)to obtain the context information of sequences both from forward and backward directions to predict the various molecular properties(e.g.,toxicity and biological activity).With comparison to the baselines,our proposed method achieves the best ROC-AUC metric.(2)Research on compound-protein interaction(CPI)prediction task.Current machine learning methods for CPI prediction mainly use one-demensional of compound and protein strings or the specific descriptors.However,they often ignore the fact that molecules are es-sentially modeled by the molecular graph.To address the above limitations,we present an end-to-end deep learning framework based on molecular graph representation,named Graph CPI,which captures the structural information of compounds and leverages the chemical context of protein sequences for solving the CPI prediction task.Compared with baseline methods,our proposed Graph CPI achieves the highest AUC result.(3)Research on drug-target binding affinity(DTA)prediction task.Most of the conven-tional DTA prediction are simulation-based methods,which rely heavily on domain knowledge or the assumption of having 3D structure of the targets,which are often difficult to obtain.Meanwhile,traditional machine learning-based methods apply various features and descrip-tors,and simply depend on the similarities between drug-target pairs.To address the above limitations,we present an end-to-end deep learning framework based on molecular graph and sequence representation,called Deep GS,which uses deep neural networks to extract the local chemical context from amino acids and SMILES sequences,as well as the molecular structure from the drugs,to predict the binding affinity between drug-target pairs.With comparison to the baseline methods,our proposed Deep GS achieves the best MSE,r_m~2and AUPR results.(4)Research on drug-drug interaction(DDI)prediction task.Most of existing computa-tional models with AI techniques often concentrate on integrating multiple data sources and combining popular embedding methods together.Yet,researchers pay less attention to the po-tential correlations between drug and other entities such as targets and genes.To address the above limitations,we propose an end-to-end graph neural network based on knowledge graph,called Knowledge Graph Neural Network(KGNN),to resolve the DDI prediction.Our frame-work can effectively capture drug and its potential neighborhoods by mining their associated re-lations in KG.To extract both high-order structures and semantic relations of the KG,we learn from the neighborhoods for each entity in KG as their local receptive,and then integrate neigh-borhood information with bias from representation of the current entity to model high-order topological information,and to obtain drugs potential long-distance correlations.Compared with several baselines,our proposed KGNN achieves the highest ACC,AUPR,AUR-ROC and F1 scores.