stDSVA: spatial transcriptomics deconvolution with a semi-supervised framework and cell type variation analysis

Spatial transcriptomics (ST) integrates transcriptomic profiling with tissue architecture, enabling the investigation of spatial heterogeneity in gene expression within intact tissues. However, due to the limited resolution of many current ST platforms, computational deconvolution methods have been developed to infer cell type compositions and/or cell type-specific gene expression profiles (GEPs) from ST data. Existing deconvolution methods mainly rely on cell type gene expression signatures derived from referenced single-cell RNA sequencing (scRNA-seq) data, and often neglect the batch effects between scRNA-seq and ST platforms. Moreover, all cell types defined in the referenced scRNA-seq data are usually assumed to be present in the ST data, leading to biased deconvolution results. Besides, ST-specific deconvolution methods mostly identify cell type compositions only. To address these limitations, we propose a novel method for spatial transcriptomics deconvolution with a semi-supervised framework and cell type variation analysis, stDSVA, which jointly estimates both cell type proportions and cell type-specific GEPs. To provide reliable supervised information, stDSVA introduces pseudo-spots generated from the referenced single-cell data and integrates the pseudo- and real-spots by batch effect removal and spatial location alignment. Gene set variance analysis is incorporated to suppress interference from irrelevant cell types. stDSVA maintains the gene-wise expression similarity and spot-wise similarities calculated from gene expression and spatial location during deconvolution. We benchmark stDSVA against existing methods on both simulated and real-world datasets. Results demonstrate that stDSVA consistently achieves high accuracy and robustness in estimating cell type proportions and cell type-specific GEPs across different scenarios.