Exploring the Latent Information in Spatial Transcriptomics Data via Multi-View Graph Convolutional Network Based on Implicit Contrastive Learning

Abstract

Latest developments in spatial transcriptomics enable thoroughly profiling of gene expression while preserving tissue microenvironment. Connecting gene expression with spatial arrangement is key for precise spatial domain identification, enhancing the comprehension of tissue microenvironments and biological processes. However, accurately analyzing spatial domains with similar gene expression and histological features is still challenging. This study introduces STMIGCL, a novel framework that leverages a multi-view graph convolutional network and implicit contrastive learning. First, it creates neighbor graphs from gene expression and spatial coordinates, and then combines these with gene expression through multi-view learning to learn low-dimensional representations. To further refine the obtained low-dimensional representations, a graph contrastive learning method with contrastive enhancement in the latent space is employed, aiming to better capture critical information in the data and improve the accuracy and discriminative power of the embeddings. Finally, an attention mechanism is used to adaptively integrate different views, capturing the importance of spots in various views to obtain the final spot representation. Experimental data confirms that STMIGCL significantly enhances spatial domain recognition precision and outperforms all baseline methods in tasks such as trajectory inference and Spatially Variable Genes (SVGs) recognition.