Research And Design Of Drug-Drug Interaction Prediction Model Based On Machine Learning

Drug-drug interactions,also known as adverse drug reactions,occur when humans take multiple drugs at the same time,and multiple drugs interact with each other in the body to affect the effects of these drugs in the body,and drug-drug interactions can cause serious physical harm or even death to patients.In the context of computer-aided drug research and the accumulation of large biomedical datasets,more and more machine learning methods are being successfully used for drug-drug interaction prediction.In this paper,inspired by pharmacological effects and based on the multilevel characteristics of drugs,three machine learningbased methods are proposed to predict drug-drug interaction relationships,and to optimize the models to improve drug-drug interaction prediction performance by continuously improving the problems of existing methods.In this paper,we first propose a drug-drug interaction prediction method MLGNN based on multi-level feature fusion.The method is able to obtain information on the internal structure of drug molecules including atomic and bond features,as well as information on the external bound protein features of the drug.Then,the joint representation learning module of MLGNN can not only fuse the two feature information well but also learn the topological structure information between drug-drug interactions on the interaction network.This enables the model to take into account the correlation between drug-drug interactions and drug-related binding proteins such as targets and enzymes as well as the complementarity of drug internal and external crosslevel features,thus enabling the model to contain rich multi-level feature information for drug-drug interaction prediction,and experiments show that the method is able to learn effective feature information of drugs,and the effect on three different size datasets is always The experiments show that the MLGNN is able to learn drugeffective feature information and consistently outperforms the current best method on three different size data sets,which not only shows the effectiveness of the model but also indicates that the model has good robustness.Then an adaptive learning method DDI-Transform is proposed on top of the MLGNN,which enables the model to learn features that are beneficial to the prediction task more effectively compared to the MLGNN.The method alleviates the problem of introducing some noisy information directly on the interaction graph using graph neural networks,while adaptively selecting similar drug pairs and information associated between drug-drug interaction events,thus enabling the model to predict fine-grained drug-drug interactions,i.e.drug-drug interaction events,more accurately.Experiments on two real datasets of different sizes show that the proposed method consistently outperforms the current optimal method and has good robustness.Finally,based on the previous two methods,a graph contrast learning framework is proposed to train the model in combination with the MLGNN and the DDI-Transform,respectively,to alleviate the over-smoothing problem that exists after the introduction of graph neural networks.The experimental results show that the graph contrast learning framework proposed in this paper can further improve the effectiveness of the model.