Abstract: The adverse reactions of drug-drug interactions have become one of the important reasons for the increase in the incidence of diseases such as digestive system diseases and cardiovascular diseases,and leads to the withdrawal of drugs from the market.Therefore,accurate prediction of drug interactions attracte widespread attention.Aiming at the problem that the traditional Encoder-Decoder model cannot capture the dependence between drug substructures,this paper proposes a TransDDI(TransformerDDI) based on Transformer and LSTM drug interaction prediction model.TransDDI includes three parts:data preprocessing module,latent feature extraction module and mapping module.The data preprocessing module uses the SPM algorithm to extract the frequent substructures that characterize the drug from the SMILES format input of the drug to form the drug feature vector,and then generate the feature vector of the drug pair.The latent feature extraction module uses Transformer to fully mine the information contained in the substructures of the feature vector,highlight the different important roles of different substructures,and generate potential feature vectors.The mapping module mainly forms a dictionary representation of the potential feature vector of the drug pair and the vector of the frequent substructure in the database,and uses the neural network fused with LSTM to make predictions.Onreal data sets BIOSNAP and DrugBank,the proposed method is compared with other 6 machine learning and deep learning methods by experiments.The results show that TransDDI has a higher accuracy rate and is convenient for drug interaction prediction.

Key words: Deep learning, Drug-drug interaction, Long short-term memory, Transformer