Clustering Method for Spatial Transcriptomics Data Based on Spectral Embedding

Spatial transcriptomics technology enables researchers to acquire both spatial location information and gene expression data within tissues, providing new perspectives for understanding tissue structure, developmental processes, and disease mechanisms. However, spatial transcriptomics data exhibits high dimensionality and sparsity, posing challenges for efficient and accurate cell clustering. To address this limitation, we propose an improved clustering method based on the CCST algorithm. Using adjustable weighting coefficients, the method integratively fuses spatial coordinates with gene expression features to enhance its perception of both spatial organization and molecular profiles. It employs graph convolutional network and deep graph infomax for feature representation learning. Spectral embedding is then applied for dimensionality reduction, followed by clustering with the k-means++ algorithm. Experiments on the 10x Visium spatial transcriptomics dataset demonstrate that the proposed approach outperforms CCST in clustering accuracy, quality, intra-cluster compactness, inter-cluster separation, and computational efficiency, while exhibiting superior stability.