Deconvolution Of Spatial Transcriptome Data Via Variational Graph Auto Encoder

Single-cell transcriptome(sc RNA-seq)sequencing can represent transcriptomic maps at the resolution of individual cells,while spatial transcriptomics,as an extension of single-cell transcriptomics,can link the spatial information of different points in tissue sections with the RNA abundance of cells within each point,which is particularly important for understanding tissue cell structure and function.However,for spatial transcriptome data,since the sampling points of the spatial transcriptome often contain multiple single cells,the gene expression measured at each sampling point is a mixture of cells with different cell types.Therefore,we use deep learning methods to reveal the cellular composition of each sampling point in the spatial transcriptome.In this study,our main work is as follows:In terms of data extraction,the initial single-cell gene expression data is pre-trained and the highly variable genes of the spatial transcriptome are screened,the low-expression or ubiquitous genes are filtered out,and then the spatial transcriptome data features are extracted by autoencoder and the spatial transcriptome data input is used to generate a data link map to learn the local features in the input more effectively.Through the variational graph autoencoder-based spatial transcriptomics data deconvolution method,we combine variational graph autoencoder with mathematical models to accurately deconvolve and restore the cellular composition of the gene expression observed at each point,thereby achieving high-level segmentation and revealing the structure of cellular heterogeneity within spatial tissues.Introduce multiple constraints into the objective function of the learning framework and iteratively optimize the analysis results.In addition,the autoencoder reconstruction loss is added to the objective function to find an embedding space more suitable for variational autograph encoder learning.In this thesis,the results obtained by the above training are combined with the regularized soft K-means to construct the spatial transcriptome deconvolution algorithm model CAVGAE.To illustrate the performance of the model framework,the deconvolution effect was verified on multiple datasets,and our method not only demonstrated superior performance on synthetic spatial data generated by different protocols,but also efficiently identified the spatial composition of cells in mouse cortical and intestinal cancer tissues.In conclusion,our method accurately reveals cellular state and subpopulation distribution based on spatial localization.In terms of results,we performed well in judging various indicators of the proportion of cells at the spatial location of the synthetic data,which was better than the results of existing spatial transcriptome analysis data methods.The performance on real data is in line with the biological law of cell distribution.Finally,this thesis summarizes the research work of spatial transcriptome deconvolution and looks forward to future work.