Generative Self-Supervised Graphs Enhance Integration, Imputation and Domains Identification of Spatial Transcriptomics

Abstract Recent advances in spatial transcriptomics (ST) have opened new avenues for preserving spatial information while measuring gene expression. However, the challenge of seamlessly integrating this data into accurate and transferable representation persists. Here, we introduce a generative self-supervised graph (GSG) learning framework to accomplish an effective joint embedding of spatial locations and gene expressions within ST data. Our approach surpasses existing methods in identifying spatial domains within the human dorsolateral prefrontal cortex. Moreover, it offers reliable analyses across various techniques, including Stereo-seq, Slide-seq, and seqFISH, irrespective of spatial resolution. Furthermore, GSG addresses dropout defects, enhancing gene expression by smoothing spatial patterns and extracting critical features, reducing batch effects, and enabling the integration of disparate datasets. Additionally, we performed spatial transcriptomic analysis on fetal human hearts, and effectively extracted biological insights using GSG. These experiments highlight GSG's accuracy in identifying spatial domains, uncovering specific APCDD1 expression in fetal endocardium, and implicating its role in congenital heart disease. Our results showcase GSG's superiority and underscore its valuable contributions to advancing spatial-omics analysis.