Identification Of Adverse Drug Reactions And Prediction Of Their Frequency Based On Deep Learning

Adverse drug reactions(ADRs)are a major issue to be addressed by the pharmaceutical industry.Among them,the frequency of adverse drug reactions is also a key factor in the risk-benefit assessment of the drug development process.Early and accurate detection of potential adverse reactions and the frequency of adverse reactions can help improve drug safety and reduce financial outlay.In recent years,computer-aided approaches have been extensively applied in the biomedical domain,which have the characteristics of low cost and short cycle compared with the traditional experimental methods.The main research of this paper is the prediction of adverse drug reactions and their frequency,and the specific research work is as follows:1.By introducing a "multi-level feature fusion deep learning model",a new prediction model,i ADRGSE is developed for the identification of adverse drug reactions in drug discovery.i ADRGSE integrates a self-attention module and a graph network module to extract the one-dimensional substructure sequence information and two-dimensional chemical structure graph information of drug molecules.In order to show the advantages of the model feature extraction methods,this paper compares i ADRGSE with several other classical methods on the ADRECS dataset for experiments,and the validation method adopts the five-fold cross-validation and the independent test set.The experimental results show the significant contribution of sequence features and graph network features to the improvement of model performance.Finally,i ADRGSE is validated on the OMOP adverse drug reaction benchmark dataset using the leave-one-out method.All experiments show that i ADRGSE outperforms the existing state-of-the-art prediction methods.For the convenience of medical practitioners and researchers,we have released a user-friendly website http://121.36.221.79/i ADRGSE,which can be used for other adverse drug reaction predictions.In addition,The source application and dataset can be downloaded at https://github.com/cathrienli/i ADRGSE.2.A new multi-task learning framework,i ADRMST,is developed by fusing drug multi-source information features and adverse reaction features through deep learning,which can be used for adverse drug reaction and frequency identification.The drug features are sequence features based on Transformer’s self-attention,graph features based on graph isomorphic network GIN,compound-compound interaction scores based on STITCH database and substructure similarity features,respectively.The adverse reaction features are obtained from Bio BERT,a pre-trained model based on biomedical corpus.i ADRMST integrates drug molecule and adverse reaction features,and determines the number of layers and neurons of the deep learning network DNN and the optimal hyperparameters of the model by five-fold cross-validation.Compared with existing forecasting approaches,experimental results show that the i ADRMST predictor has more robust and stable performance.Researchers can access the prediction procedure of this study at https://github.com/cathrienli/i ADRMST.