Abstract
Recent advancements in Spatial Transcriptomics (ST) technologies have enabled researchers to investigate the relationships between cells while simultaneously considering their spatial locations within tissue. These technologies facilitate the integration of gene expression data with spatial information for clustering analysis. While many clustering methods have been developed, they typically rely on the dataset's intrinsic features without incorporating domain knowledge, such as marker genes. We argue that incorporating marker gene information can enhance the learning of cell embedding and improve clustering outcomes. In this paper, we introduce MGGNN (Marker Gene-Guided Graph Neural Networks), a novel approach designed to enhance spatial transcriptomics clustering. Firstly, we train the model using a contrastive learning framework based on a Graph Neural Network (GNN). Subsequently, we fine-tune the model using a few spots labeled by the expression of marker genes. Simulation and experiments conducted on two real-world datasets demonstrate the superior performance of our model over state-of-the-art methods.