Anticancer Drug Response Prediction Based On Graph Neural Network

The in-depth study of a large amount of cancer genome data has found that even among patients with the same cancer,there are large differences in individual gene mutations.Using patient-specific genomic profiles to design personalized treatment plans for patients can play a key role in controlling the progression of the disease.Therefore,accurate prediction of anticancer drug response is an important part of personalized therapy today.Traditional drug response assays require a large number of clinical trials,with huge investment costs and low execution efficiency.Using computer-related knowledge to construct algorithmic models to predict the response of anticancer drugs has important practical significance for preclinical prediction.In this paper,two different anticancer drug response prediction models were constructed using deep learning algorithms and graph neural networks,combined with multi-omics features of cancer cell lines and drug feature information.Its main work is as follows:1)Considering the heterogeneity of network nodes in the drug-cell line response prediction problem,a heterogeneous graph convolutional network-based anticancer drug response prediction method HGCNDRP is proposed.Based on the feature data of drug structure fingerprints,gene expression profile data of cell lines and known drug-cell line interaction data,the algorithm combines heterogeneous graph convolutional networks and deep neural networks.Firstly,the drug similarity and cell line similarity are calculated separately to construct a heterogeneous network;secondly,the feature representation of the drug and cell line is learned through the heterogeneous graph convolution model;after that,the DNN model is used to predict the drug response of the cancer cell line.Finally,the 5-fold cross-validation method and other aspects are tested in the GDSC data set,and compared with HRWR,HNMDRP,TMF and NRL2 DRP algorithms.The AUC value of HGCNDRP is 20.97% higher than HNMDRP,10.41% higher than HRWR,16.06% higher than NRL2 DRP,and 12.32% higher than TMF.Overall,HGCNDRP has better prediction effect,better performance.2)The emergence of genetic multi-omics data information can help people understand biological processes more comprehensively,and is also of great significance to the prediction of drug response.Therefore,based on gene multi-omics data and drug molecular map data,a drug response prediction method MBCDR based on multi-branch network is proposed.This method adopts different characterization methods for the characteristics of different branches,and for cell lines,its three genomic characteristics are characterized separately.For drugs,starting from the molecular graph of drugs,a graph isomorphic network is used to capture and update representations of atomic nodes,and to connect representations at different levels,so as to be able to fuse information of different scales.Finally,the cell line characterization and drug characterization are assembled,and the fully connected neural network is used for drug response prediction.Based on the GDSC dataset,5-fold cross-validation and independent verification are performed.In terms of AUC scores,compared with the three best baselines of Graph CDR,Deep CDR and NRL2 DRP,they exceeded 1.43%,3.45% and 4.70%,respectively.In terms of AUPR scores,they were 4.26%,8.65% and 12.51% higher respectively.The Accuracy indicators are 2.21%,1.16% and 0.89%higher respectively.The F1＿Scores indicators have increased by 2.09%,4.05% and 9.59%respectively.The cross-validation results show that the prediction performance of MBCDR is better.Under independent test verification,MBCDR also achieved better classification results.