A deep learning representation and spatial Bayesian cell-type deconvolution for spatial transcriptomics
Abstract
Background. Spatial transcriptomics provides unprecedented insights into gene expression in the spatial context of tissues; however, some mainstream techniques lack quantitative analysis of uncertainty.
Methods. In this study, we propose a novel computational framework that combines deep Gaussian processes (DGP) with Bayesian uncertainty for spatially aware deconvolution analysis of spatial transcriptomics data. Our method leverages the hierarchical non-linear mapping capabilities of DGP to capture complex spatial dependencies while maintaining probabilistic interpretability through Markov Chain Monte Carlo (MCMC) inference. Our approach introduces several key innovations: (1) Use a multi-layer hierarchical structure, with each layer being a Gaussian process capable of learning nonlinear mappings. (2) In a neural network, forward propagation uses the output of the previous layer as the input for the next layer. and (3) Using variational inference to approximate the posterior distribution.
Results. We compared our method with existing methods using a pancreatic ductal adenocarcinoma dataset containing matched single-cell RNA-seq and spatial transcriptomics data. Our method demonstrated greater accuracy in estimating cell type proportions (especially rare cell types) than other existing methods and revealed biologically meaningful spatial patterns of T cell distribution. The Bayesian nature of the method provides uncertainty quantification for each estimate, enhancing the interpretation of spatial heterogeneity in complex tissues.