Research On Deep Learning-based Algorithms For Drug-drug Interaction Prediction

With the rapid development of biopharmaceutical technology,Drug-Drug Interaction(DDI)has become an important part of drug research in the medical field.Drug-Drug Interaction is when two or more drugs are used together and one can affect the efficacy of the other,leading to adverse effects,antagonism or synergy.Predicting DDI is therefore an issue that clinicians and drug developers need to focus on.Many computational methods and tools are now available for predicting DDI,including methods based on structure,drug metabolic pathways and drug-target interaction.These methods can improve the accuracy and efficiency of DDI prediction and play an important role in new drug development and clinical applications.However,predicting DDI remains a challenging problem that requires further research and improvement.To address the problems that most existing drug interaction prediction algorithms fail to integrate drug molecule features and drug interaction network structures well,this paper proposes a DDI prediction algorithm DMFDDI based on deep multimodal feature fusion,which fuses multiple features such as drug molecule graphs,drug interaction networks and biochemical feature similarities of drugs to predict DDI.In the drug molecule structure extraction module,an attention gated graph neural network is introduced to obtain global features of the molecule graph and local features of each atom to fully extract the drug molecule structure.In the drug interaction network feature extraction module,a sparse graph convolutional network is introduced to learn the topological structure information of the DDI network.In the multimodal feature fusion module,an attention mechanism is used to efficiently fuse different features.To verify the performance of DMFDDI,this paper compares it with eight current state-of-the-art algorithms.The results of the comparison experiments show that DMFDDI has better performance in DDI prediction.The impact of different modules of the algorithm on performance is also analyzed through ablation experiments.The interpretability of DMFDDI prediction results is verified by case study.To address the problem that existing drug interaction prediction algorithms are less involved in multi-type prediction,this thesis proposes a multi-type drug interaction prediction algorithm MGDF based on multi-graph convolution as well as feature depth fusion.To investigate the prediction of multi-type drug interactions,a nearest neighbour view is first constructed based on feature similarity,and a diffusion view is added in addition to the neighbour view of local topological relationships.The integration of multiple views helps capture more comprehensive drug features;meanwhile,a Multi-Head attention mechanism and residual connectivity are added to the graph convolution module to improve the model;finally,feature extraction and fusion are performed in the feature extraction module and the feature fusion module,respectively,introducing an attention mechanism to effectively fuse different drug features to more comprehensively characterize the drugs and improve the richness of drug embedding information.To validate the performance of MGDF,this paper compares it with six current state-of-the-art methods.The results of the comparison experiments show that MGDF has better performance in DDI multi-type drug interaction prediction,and the contribution of each module of the MGDF model to the prediction results is verified by ablation study.