SpaViT: Self-supervised Prediction of High-Resolution Spatial Transcriptomics with Vision Transformer

The spatiotemporal specificity of gene expression highlights the importance of integrating cellular spatial information to better understand the specific functions of cells within tissues. The advent of spatial transcriptomics (ST) technologies has made it possible to quantitatively measure gene expression in cells while also pinpointing their exact locations within tissues. However, widely used ST techniques are frequently limited by low resolution, potentially hindering researchers from fully understanding gene expression patterns, cell type distribution, and their interactions. Here we propose SpaViT, a self-supervised method based on the Transformer architecture for predicting high-resolution gene expression. SpaViT leverages customized self-supervised proxy tasks to learn the continuous patterns of gene expression within tissues and predicting high-resolution gene expression profiles.We evaluate the performance of SpaViT on diverse datasets from different platform swith different technologies. The results indicate superior performance of SpaViT in enhancing spatial resolution and predicting gene expression in unmeasured areas compared to other deep learning and traditional interpolation methods. Additionally, SpaViT enhances the spatial patterns of gene expression, aiding researchers in identifying biologically significant differentially expressed genes and pathways. Our source code and all datasets used in this study are available at https://github.com/wenwenmin/SpaViT and https://zenodo.org/records/14160324.