Prediction Research Of Drug-drug Interaction Based On Multi-graph Data Fusion

Drug combination therapy is a highly promising strategy to overcome drug resistance and improve the efficacy of monotherapy and has been proven to reduce dose related toxicity.But apart from inter drug synergistic reactions,there are some antagonistic drug interactions,which are the main reasons for producing adverse drug events(ADEs).Most current research methods mainly focus on the dichotomous problem,that is,predicting whether there is an interaction between two drugs,and the problem of multiclassification has less to do with.Accurate prediction of the type of drug-drug interactions(DDIs)is important for both drug development and effective application of drug combination therapies.In this paper,deep learning algorithms are applied,combined with the information of drug characteristics,to construct two different prediction models of drug drug interactions.Its main work is as follows:(1)Most of the currently proposed methods are based on only one or a few features of drugs and do not consider the topological relationships of feature information networks.To integrate multiple features of drugs more effectively,NMDADNN,a drug-drug interaction prediction method based on multimodal data fusion,was constructed based on similarity assumption.Drug interaction events and five drug features were first collected from the Drug Bank dataset to construct a similarity network using the Jaccard coefficient.Second,a random walk algorithm with restart was employed to capture the topological information of each similarity network,and positive point mutual information was calculated to obtain the topological similarity between nodes.After that,a multimodal deep autoencoder(MDA)was employed,fusing five drug similarity matrices to obtain a uniform feature description of drugs.Finally,the types of DDIs was predicted using the DNN model.Compared with two other DNN based methods,DDIMDLand Deep DDI,in a 5-CV test on the Drug Bank dataset.The ACC,AUPR,AUC,F1 score,Precision and Recall indexes of NMDADNN were 6.1%,6.9%,0.3%,12.2%,14.7% and 12.0% higher than those of Deep DDI,and 1.3%,3.8%,0.05%,4.8%,2.8%and 6.3% higher than those of DDIMD,which achieved better prediction results.(2)Similarity calculations can be disturbed by noise or outliers,leading to inaccurate similarity calculations.And knowledge graph,which fuse information from different data sources,can improve data accuracy and completeness.In the graph domain,models such as the graph attention network(GAT)can automatically learn feature representations of graph structures.Fusegat,a subtomogram learning model based on knowledge maps and attention mechanisms,is proposed for more efficient feature extraction of drugs.A knowledge map is first constructed based on the collected experimental data,and then the initial features of entities and relationships are obtained using the knowledge map method,Transe,with initial operation for all entities embedding.For a given drug pair,a local subgraph associated with the drug pair is selected from the knowledge map.Then,features of subgraphs are generated with attention mechanisms and message passing mechanisms.Finally subgraphs and drug pair features are integrated and a decoding classifier is used to predict the type of drug pair.Fusegat’s F1,ACC,Cohen’s kappa metrics on the Drug Bank dataset increased the predictive performance effectively by 9.65%,2.4%,2.15% compared to the best model with other contrasts.