Researching On Prediction And Application Of Drug-drug Interactions Based On Learning With Graph-structured Representations

Drug-drug interaction is an important consideration in drug development,clinical application,effective intervention and promotion of life process.Accurate identification of drug-drug interactions can not only avoid many medical accidents but also promote the development of pharmacogenomics,pharmacometabonomics,pharmacotranscriptomics and pharmacoproteomics.However,the traditional biotechnology experiment is very expensive and time-consuming.Therefore,it is urgent to use large-scale data calculation and prediction methods as decision-making aid to pre-screen candidate drugs with high confidence to alleviate the above problems.In this paper,the following work has been done on how to effectively use computer technology and machine learning technology to predict potential drug interactions:(1)In order to predict the potential drug-drug interactions more accurately and robustly,a new prediction framework is improved.The framework mainly consists of three parts: drug interaction topological structure information extraction module,drug chemical sequence information extraction module with chemical significance and drug biological function information extraction module with biological significance.On this basis,the information of the three modules is fully fused by a self-attention mechanism or other methods to generate a highly representative comprehensive feature descriptor,and the final prediction is made by Deep Neural Network(DNN).Based on the framework,this thesis proposes three prediction models by using feature extraction algorithms from different angles(such as homogenous graph learning and heterogeneous graph learning).A large number of experimental results show that the performance of these three models is better than other comparison models.(2)With the research on drugs,a large number of computational models have been developed.However,most models did not pay much attention to the rich side information of the drug knowledge graph.From the perspective of graph structure representation,most drugdrug interaction prediction models used common graph embedding methods,which ignored other heterogeneous knowledge information of drugs,such as heterogeneous node information and heterogeneous edge information of drug knowledge graphs.Therefore,a new model of CCSDDI is proposed to solve the above problems.In this model,firstly,the knowledge graph embedding(KGE)algorithm is used to fully extract the feature information of various nodes and edges in the knowledge graph,then the word embedding algorithm of CBOW and Similarity Network Fusion(SNF)algorithm are used to learn pharmaceutical chemical information and pharmaceutical biological information,and finally,the selfattention mechanism is used to fully fuse various features,and the multi-layer perception is used to make predictions.The test results in common public data sets prove that our method has good performance,and the advantages of our method are reflected in the comparative test.The ablation experiment proves the importance of all modules.Finally,in the case study of Malignancies-related drugs,it is proved that the proposed model can also be applied to real life with reliable prediction results.(3)In recent years,many graph embedding methods have been proposed to extract graph structure information and have been applied to solve bioinformatics problems.However,most models did not consider and analyze the relation of graphs and subgraphs,so there are limitations in the task of global information learning.Therefore,a new model SHBDDI is proposed.It uses iterative coarse-grained learning based on graph collapse to obtain stronger feature vectors.We use the method of Hierarchical Representation Learning for Networks(HARP)to extract graph structure features.Additionally,we propose a novel method based on a graph neural network to learn biological function information.It can aggregate and diffuse features based on different functional spaces,and finally integrate feature information of all functions to obtain embedded vectors of all entities.For pharmaceutical chemical information,we use the word embedding method of Skip-gram to calculate the digital representation of the drug substructure sequence.Finally,the self-attention mechanism is used to fully fuse all kinds of features and make a prediction.In the experiment,the ablation experiment verifies the contribution level of each module,and a series of comparison experiments of feature extraction and classifier prove the advantages of our model.In order to apply the model to reality,we made a case study of related drugs in COVID-19 and gave a set of candidate drugs with high confidence.It provides new insights and helps with the prediction of potential drug interactions.(4)With the increase in the number of approved drugs every year,the number of unknown drug interactions continued to rise.Many models have been proposed to predict unknown drug-drug interactions.However,these methods often have a single consideration angle and poor inductive ability.Therefore,based on the biochemical characteristic information of drugs and the Laplacian Eigenmaps(Lap)algorithm,we proposed a new computational model,which is called Tri SDDI.The model fully considers drug characteristic information from various angles.On the basis of extracting drug biochemical information by using Skip-gram and SNF methods,the degree information of all nodes in the drug-drug interaction network is integrated into the model,and the Logistic Regression Classifier(LRC)is used to predict the results,which can improve the prediction ability of the model.