Research On Drug Repositioning With Graph Information Fusion And Heterogeneous Networks

The process of developing new drugs is expensive and time-consuming,and has a high attrition rate.Improving the success rate of drug development can help to significantly reduce the cost of drug discovery.Drug repositioning technology,i.e.,identifying new applicable symptoms from approved or mature clinically used drugs,can effectively improve the efficiency of drug discovery.Therefore,drug repositioning plays a crucial role and has become a hot research topic in drug discovery.The accumulation of biomedical data has provided rich data for drug repositioning research,but how to mine the information in drug molecular structure maps and drug-disease associations in multi-source data has always been an important issue of concern for researchers in this field.To this end,in this thesis,based on graph convolutional neural networks,the following two research works are conducted.(1)To address the problem that the molecular structures of drug compounds are not fully explored in existing drug-target interaction prediction tasks,this article proposes Transformer Network Incorporating Multilayer Graph Information(Deep MGT)for drug-target interaction prediction.In Deep MGT,the SMILES sequences of drug molecules are represented as graph structures.This model utilizes a transformer network incorporating multilayer graph information,which captures the features of a drug’s molecular structure so that the interactions between atoms of drug compounds can be explored more deeply.At the same time,a convolutional neural network is employed to capture the local residue information in the target sequence,and effectively extract the feature information of the target.The experiments on the Drug Bank dataset showed that the Deep MGT can be effective for DTI prediction and outperforms models based on the structure of target sequences.The drug repositioning experiment on COVID-19 demonstrated the proposed Deep MGT ability to find therapeutic drugs in drug discovery.(2)In view of the drug repositioning research methods lack representation of neighborhood information and rely on prior knowledge,resulting in weak generalization ability.This thesis proposes the Heterogenous Neighborhood Networks with Contrastive Learning(HNNCL)for drug repositioning research.In HNNCL,single-neighborhood feature extraction module and multi-neighborhood feature extraction module based on graph convolutional neural network are designed to extract the feature information of drugs and diseases in drug-drug(disease-disease)similarity network and drug-disease association network,respectively.The single domain and multi neighborhood features are fused and optimized by using comparative learning and ternary loss,enhancing the information interactions of different neighborhood spaces,and attention mechanism is used to integrate the multi-source information finally.Experimental results on multiple datasets confirm that the performance of HNNCL is better than several current state-of-the-art prediction methods.It can effectively perform drug-disease association prediction.HNNCL was demonstrated to have better ability to discover therapeutic drugs in drug repositioning experiments for Alzheimer’s disease.The Deep MGT and HNNCL methods proposed in this thesis can effectively solve the related problems in drug repositioning research with good performance,which provides a new method for drug repositioning.In the future,we will take the spatial structure of targets and proteins as the basis.In order to improve the practical value of the model and reduce the cost of drug development,we are committed to using the spatial structure information of targets and proteins for drug repositioning research.