Deep Learning Based Method For Predicting Drug-Drug Interactions Events

Millions of patients die from Drug-Drug Interactions(DDIs)each year and therefore DDIs have attracted widespread attention.A large number of deep learning-based methods have been used to testify whether there is an interaction between drugs in previous studies.The current deep learning methods mainly have three shortcomings.First,deep learning methods generally suffer from insufficient data,and reliable drug-drug interaction data is very scarce.Previous studies were based on only a small amount of drug-drug interaction data and did not take full advantage of unlabeled drug data.These unlabeled drug data are readily available and contain semantic and structural information about the drug,which can help drug-drug interactions prediction.This article will examine how to effectively utilize unlabeled drug data.Second,the current research on drug-drug interactions is limited to predicting whether there is an interaction between drugs.However,few researchers have paid attention to the prediction of drug-drug interactions events,which is more practical.This article will examine the prediction of drug-drug interactions events.Third,the current research methods mainly focus on the prediction of single features of drugs,such as chemical features of drugs,graph structure features,etc.Drug multi-type feature fusion has not been explored.This paper will study the multi-type feature fusion of drugs and verify its effectiveness through ablation experiments.This paper proposes TBPM-DDIE(Transformer Based Pretrained Methods for improving the prediction of Drug-Drug Interactions Events)model and MFFM-DDIE(Multi-type Feature Fusion Method for predicting Drug-Drug Interactions Events)model for the above three problems.The TBPM-DDIE model makes full use of unlabeled drug data through the Transformer-based pre-training method,and obtains the feature vector encoding with drug structure and semantic information through the trained Transformer encoder.Then,the Pretrained feature and the Chemical feature of the drug are directly concatenated,and the fused features are input into the classifier of the fully connected neural network to predict the Drug-Drug Interactions Events.The MFFM-DDIE model proposed a GNN-based drug molecular graph structure information encoding model,and the Graph feature are obtained by encoding the graph structure information.Then MFFM-DDIE conducted a multi-type feature fusion method on Chemical feature,Pretrained feature and Graph feature of drugs.Three multi-feature fusion methods are proposed,direct concatenate,merged attention and co-attention.And the effectiveness of multi-type feature fusion for Drug-Drug Interactions Events task is verified by ablation experiments.We have done experiments on the real-world dataset and compared it with the latest models.The results show that TBPM-DDIE and MFFM-DDIE can achieve state-of-theart effects on predicting DDIs events.