Prediction Of Drug-Drug Interaction Based On Deep Learning

With the continuous development of biomedical big data technology,huge amounts of drug molecule structures are constantly being introduced into the database.The prediction of drug-drug interactions(DDIs)is considered to be a hot topic in current drug research and plays a pivotal role in the development and repositioning of new drugs.Traditional bio-experimental methods for predicting DDIs are time-consuming and costly,while using deep learning to predict DDIs is currently an important task in bioinformatics.This paper is based on a deep learning approach to investigate drug-drug interactions.The main studies are as follows:1.A new method for drug-drug interaction prediction,SA-Bi LSTM,is proposed based on double bidirectional long short-term memory(Bi LSTM)and self-attention mechanism.First,the characteristic information of the drug is extracted using FP3 fingerprints,MACCS fingerprints,Pubchem fingerprints,and Pa DEL molecular descriptors.Second,least absolute shrinkage and selection operator(Lasso)is applied to eliminate redundant features.Then,repeated edited nearest neighbors(RENN)method is used to balance the data to get the best feature vector.Finally,the optimal feature vectors are fed into the classifier combining self attentive mechanism and Bi LSTM to predict the DDI.SA-Bi LSTM achieves high prediction accuracy on both datasets based on 5-fold cross-validation and comparison with other prediction methods.To further evaluate the predictive performance of SA-Bi LSTM,the drug-drug interaction network is validated.It is shown that SA-Bi LSTM obtains superior forecast results,which can provide a new idea for predicting DDI.2.A new method for predicting drug-drug interactions,DBGRU-SE,is proposed based on double bidirectional gated recurrent unit(Bi GRU)and squeeze-and-excitation(SE)attention mechanism.First,the characteristic information of the drug is extracted using FP3 fingerprints,MACCS fingerprints,Pubchem fingerprints,and 1D and 2D molecular descriptors.Second,Group Lasso is applied to eliminate redundant features.Then,SMOTE-ENN is used to equilibrate the dataset to acquire the optimal feature vectors.Finally,the optimal feature vectors are sent into double Bi GRU with SE attention,and the final features are fed into three dense layers to complete the prediction of DDIs.DBGRU-SE model achieved ACC values of 97.51% and 94.98% and AUC values of99.60% and 98.85% on two datasets with five-fold cross-validation,respectively.Compared with other prediction methods,DBGRU-SE achieves better performance.To further evaluate the predictive performance of DBGRU-SE,the drug-drug interaction network is validated.The validation results suggest that DBGRU-SE can significantly increase its ability to predict DDIs,which can provide a useful theoretical foundation for novel drug discovery and development.