Tissue identification on spatial transcriptomics using variational graph auto-encoder

In recent years, the rapid development of spatial transcriptomics has expanded the analysis of gen e expression patterns by adding spatial context to the traditional single-cell-level measurements. This advancement provides rich spatial information for deciphering gene expression patterns in tissues. However, how to combine gene and spatial information has become an unavoidable challenge. Many existing methods either fail to leverage spatial information adequately or inadvertently distort gene expression features when attempting to integrate spatial data. To address these challenges, we developed a framework called STVGAE, based on a Variational Graph Auto-Encoder (VGAE). STVGAE's unique strength lies in its generative modeling capability, which accurately captures the latent distribution characteristics of gene expression data. Additionally, multiple embedding paths complement each other, mitigating the potential distortion of gene features due to spatial information fitting. Validation on spatial transcriptomics data from various tissues demonstrates that STVGAE achieves high accuracy in tissue identification and maintains the integrity of gene features, thus reducing the likelihood of fragmented predictions.