Abstract
Spatial transcriptomics is able to acquire cellular gene expression while retaining spatial location. It is often accompanied by matched hematoxylin and eosin-stained histology whole-slide images. This retention of spatial information is critical for studying key issues in cell biology, developmental biology, neurobiology, and tumor biology. However, conventional sequencing technologies are costly and time-consuming, limiting the development of spatial transcriptomics research. Recently, deep learning methods have been widely applied to spatial transcriptomics prediction, but there are some problems in existing methods. To tackle the problem, we develop Reg2ST, a deep learning model to learn potential patterns in gene expression and apply them for spatial transcriptomics prediction. Reg2ST treats spatial transcriptomics and histology as different expressions of the same data. Contrastive learning is used to minimize the distance between them. Then image features are used to predict gene features, which aligns histology images with spatial transcriptomics. Reg2ST uses a novel way to capture relationships among spots instead of K-Nearest-Neighbors. Evaluations of Pearson correlation coefficient, statistical tests, computational efficiency using human breast cancer, and cutaneous squamous cell carcinoma datasets demonstrate the superior performance of Reg2ST for spatial gene expression prediction.