| With the rapid development of deep learning,more and more scholars have paid attention to the effectiveness of deep learning in exploring the laws of biological information in complex data.For example,researchers used a variety of deep learning methods combined with clustering algorithms to extract genetic features from tumor single-cell RNA sequencing(sc RNA-seq)data to achieve accurate identification of tumor subtypes.In addition,in the treatment of complex diseases such as cancer,single drug therapy has the disadvantages of easy drug resistance,high cytotoxicity and single targeting.Therefore,researchers try to use deep learning methods to predict drug combination,so as to provide candidate drug combinations for clinical treatment in a more accurate,efficient and convenient way.However,in the subject of tumor subtype identification,researchers are often trapped in the sparse and zero-inflated characteristics of sc RNA-seq data,and it is difficult for traditional clustering algorithms to obtain accurate tumor subtype identification results.Although some deep learning algorithms have used graph neural networks to effectively obtain the genetic features of cells for subtype recognition,these methods usually focus on learning the feature representation of the data itself,ignoring the structural relationship between data samples.d universities.In the subject of drug combination prediction,with the release of large data sets such as NCI-ALMANAC and Drug Comb,the advantages of deep learning in this subject have gradually emerged,and many advanced drug combination prediction algorithms based on deep neural networks and graph neural networks have also emerged in the field.However,these algorithms still have some practical problems that need to be improved to achieve the use of the topology between data,but also to reduce the computational complexity as much as possible to obtain structured global features.1)Aiming at the problem of tumor subtype identification,this paper proposes a deep structured clustering model sc DSC based on single cell RNA-seq data.The sc DSC first uses the autoencoder module based on the ZINB model to learn the embedding features of the hidden layer.Then,sc DSC uses graph neural network and autoencoder to capture the structured information of data.Finally,sc DSC uses multiple mutual supervision strategies to let different modules learn from each other,thus realizing the synchronous update of autoencoder network and graph neural network under the same framework.The experimental results show that the clustering performance of sc DSC on six real sc RNA-seq datasets is better than the five most advanced competitive algorithms.In addition,the ablation experiment analyzed the influence of each sub-module of sc DSC on the model performance in detail.The clustering visualization results show that sc DSC restores the cell data structure that the original data cannot represent well,showing a good trajectory of cell development.2)Aiming at the problem of drug combination prediction,this paper proposes a new method for combined prediction of anticancer drugs based on deep learning.DGSSyn ADR is mainly composed of three parts : constructing multivariate heterogeneous graph,drug and protein feature extraction based on LRGA and bilinear predictor.Firstly,the multivariate isomerism diagrams of drug-drug,drug-protein and protein-protein were constructed by DGSSyn ADR.Subsequently,DGSSyn ADR uses LRGA-based global weighted aggregation of nodes on the heterogeneous graph to extract feature representations of drugs and proteins in the heterogeneous graph through a multi-layer convolutional layer and an LRGA-based attention pooling layer.Finally,DGSSyn ADR uses a bilinear predictor to obtain the synergy score of the drug pair for the drug pair characteristics and cell line characteristics.On the Drug Comb dataset,the performance of DGSSyn ADR was compared with two advanced deep learning methods and five classical machine learning methods.It was verified that the prediction performance of DGSSyn ADR was significantly better than other methods.The effects of cell line and tissue specificity on model prediction performance were also analyzed.In addition,the contribution of each module of the DGSSyn ADR model to the prediction results was verified by ablation studies.Finally,through case analysis,it was proved that the combination of anticancer synergistic drugs predicted by DGSSyn ADR was consistent with many previous studies.Detailed experimental data show that DGSSyn ADR is an effective drug combination prediction tool,which will contribute to the discovery of new anticancer synergistic drug combinations. |
