Research On Single-cell Identification Method Based On Graph Neural Network

Single-cell transcriptomics is an essential component of modern systems biology,increasingly used in biomedical research,and can provide new insights into complex cellular ecosystems and underlying molecular interconnections.Single-cell RNA sequencing technology(sc RNA-seq)allows the measurement of transcriptomes of thousands of single cells from multiple biological samples under different conditions,whose expression levels can reflect the overall characteristics of the cells.Single-cell transcriptomics is now widely used to address cellular heterogeneity.With the rapid accumulation of single-cell transcriptomic data in public databases,effective and efficient cell identification methods that can leverage existing annotations to curate newly sequenced cells become critical.At the same time,the thresholds of traditional cell type identification methods for classification are usually based on experience,which lacks stability and interpretability.A graph neural network is a neural network that can model and analyze graph data,and can learn latent representations of nodes by using graph node features and structural information.Multiview learning can learn to obtain better feature representations by exploiting the complementarity between multiple views.Single-cell transcriptomes are a type of data that are well suited to be constructed into graph structures,but due to the intrinsic heterogeneity between cell clusters and extrinsic differences between datasets,different graph construction methods show obvious differences on different recognition tasks.So in this paper,we propose a heterogeneous graph neural network based cell type identification method(sc MGCN).First,the single-cell transcriptome was constructed into graph using different methods,with nodes representing cells and edges representing relationships between cells.Then,the heterogeneous graph data with different kinds of edges are input into the model,and the graph convolution network is used to train and analyze the graphs obtained by different graph constructed methods,and then the output of different graph convolution is aggregated through the attention mechanism.Finally,the multi label prediction of cells is carried out through the fully connected layer to obtain the corresponding node representation,which represents the confidence prediction scores of different cell types in the form of vector.To verify the effectiveness of the model,we conduct experiments on six sets of benchmark datasets,ten sets of cross-platform datasets,and four sets of cross-species datasets.At the same time,a comparative test was done on five cell identification methods,Seurat v3,conos,scmap,sc GCN,CHETAH.sc MGCN is more stable than other methods in the experiment and shows better accuracy in cross-species and cross-platform datasets.In addition,we experimentally compare the effects of single-cell graph structure data graph neural networks.The results show that sc MGCN can effectively integrate the advantages of different graph constructed methods.We also use other heterogeneous graph neural network models for experimental comparison,and verify that our proposed model is better than other heterogeneous graph neural networks.