The Study Of Multi-Typed Drug-Drug Interaction Prediction Based On Deep Learning

With the continuous development of combination drug therapy,adverse drug-drug interactions(DDIs)have become an increasingly serious problem in the medical and health system.Therefore,many computational models have been developed to predict potential DDIs to reduce the risks of unknown adverse DDIs.In recent years,the rapid development of deep learning and the effective application of biomedical knowledge graphs(KGs)have greatly improved the predictive performance of these models.However,the problems of feature redundancy and KG noise also arise and need to be solved urgently,bringing new challenges for researchers.To address these challenges,this thesis proposed a multi-channel feature fusion model for multi-typed drug-drug interactions prediction(MCFF-MTDDI),which was mainly composed of the feature extraction module,the feature fusion module,and the classifier module.First,the drug chemical structure features,the extra label features of drug pairs,and the three types of biomedical KG-based features of drugs were extracted respectively in the feature extraction module.Then,in the feature fusion module,it began with a State Encoder that was designed to obtained initially the key KG features of target drug pairs;A Gated Recurrent Units(GRUs)-based multi-channel feature fusion approach was proposed afterwards to effectively fused these diverse features;in addition,extra label feature fusion suitable for multi-label classification task was set to assist DDI type prediction.Finally,multi-typed DDIs were predicted by inputting the final fused representations of the target drug pairs into the fully connected neural network in the classifier module.The innovation of the proposed model mainly lied in three aspects.First,we are the first to integrate the drug pairs’ extra label information into KG-based multi-typed DDI prediction;secondly,we innovatively proposed a novel KG feature learning method that was combined with two other common graph representation methods to extract more abundant KG-based information of each drug,and designed a novel State Encoder to obtain target drug pairs’ KG features which contained more key drug-related KG information with less noisy;thirdly,a GRU-based multi-channel feature fusion approach was proposed to yield more comprehensive feature information about drug pairs,effectively alleviating the problem of feature redundancy.We comprehensively assessed the performance of the model for predicting interactions of known-known drugs,known-new drugs,and new-new drugs by fivefold cross-validation with four datasets in multi-class and multi-label prediction tasks.In addition,ablation studies and case studies were further conducted,and all the results fully demonstrated the effectiveness of the proposed model.