MiRNA-drug Resistance Association Prediction Based On Bi-Level Graph Neural Networks

Most drugs exert therapeutic efficacy by targeting specific proteins,but a considerable amount of proteins are not targeted by any drug.Due to the fact that miRNAs can regulate the expression of specific genes,and the disorders of miRNA expressions are related to a variety of human diseases,the focus of drug target selection has gradually shifted to miRNAs.Studies have shown that although drugs can cause changes of miRNA expression levels to a certain extent,and then exert therapeutic efficacy on specific diseases,miRNAs can also affect drug therapeutic efficacy by regulating the expression of related genes.The changes of miRNA expression levels in patients may be the main reason for the difference of drug resistance in different individuals.Therefore,it is of great significance to study the associations between miRNA and drug resistance for the implementation of miRNA targeted therapy.Existing miRNA-drug resistance association prediction methods are difficult to predict the miRNAs or drugs with few or no associations,and these methods cannot well exploit potential information of the known miRNA-drug resistance association graph.In addition,in the prediction process of these methods,the specificities of miRNA-drug resistance association characteristics are not taking into account.Therefore,in this paper,an Attentive Bi-Level Graph Neural Networks method(ABiGNN)is proposed to predict the associations between miRNA and drug resistance.In particular,ABi-GNN applies information completion strategy to solve the problem that miRNAs or drugs without any associated information cannot be predicted in previous models.The specific process of the model is as follows: Firstly,ABi-GNN represents the known miRNA-drug resistance associations as a graph,and uses a bi-level graph convolution neural network to learn the graph embeddings of miRNAs and drugs.At the lower level,ABi-GNN uses graph convolution neural network to learn the initial representations of miRNAs and drugs on the drug molecule graphs and miRNA expression profile graph respectively,and at the higher level,graph convolution neural network is applied on the miRNA-drug resistance association graph to learn graph embeddings of miRNAs and drugs based on the initial representations learned at the lower level.Secondly,ABi-GNN combines graph embeddings of miRNAs and drugs with other biological characteristics of miRNAs and drugs to represent the features of miRNA-drug resistance association pairs,and applied attention mechanism to represent the different contributions of different features in the final prediction.Finally,deep neural network is applied to predict and score the unknown miRNA-drug resistance association pairs.Experiments show that ABi-GNN produces a great performance in miRNA-drug resistance association prediction.Through the analysis of the model structure,it can be concluded that graph embeddings of miRNAs and drugs learned by the bi-level graph neural network is very effective for the prediction task;attention mechanism also plays an important role in the feature representation of miRNA-drug resistance association pairs.Compared with the benchmarks and other state-of-the-art methods,ABi-GNN predicts the miRNA-drug resistance associations more accurately.Case studies also verify the great performance of ABi-GNN.