STCGCar: Graph Contrastive Learning with Reliable Augmentation for Spatial Transcriptomics Clustering

Accurately identifying spatial domains based on spatial transcriptomics (ST) data can greatly promote our understanding of cellular composition and tissue organization. While graph neural networks (GNNs) have shown significant advancements in spatial clustering, they tend to be insensitive to noisy edges, leading to intersections among identified spatial domains. Here, we introduce a GNN-based ST Clustering framework, called STCGCar, utilizing a Graph Contrastive learning model with reliable augmentation and redundancy reduction strategies. The framework begins by creating an enhanced view through a reversible network after data preprocessing. Subsequently, low-dimensional embeddings of spots are learned using a multi-head attention mechanism. Moreover, a redundancy reduction strategy is employed to reduce information redundancy in potential feature space. Finally, spatial domains are delineated through K-means clustering, followed by downstream analysis. STCGCar was benchmarked against six state-of-the-art clustering methods (i.e., Seurat, conST, CCST, STAGATE, DeepST, and GraphST) using five 10x Visium datasets, a STARmap dataset, and two Stereo-seq mouse embryo datasets. Through evaluation with adjusted rand index (ARI), normalized mutual information (NMI), and four internal indicators, it demonstrated outstanding clustering performance compared to other methods on four labeled and four unlabeled datasets. Additionally, STCGCar accurately identified spatial domains and discovered three potential differentially expressed genes (AZGP1, CD24, and CCND1) in human breast cancer tissues. Furthermore, it effectively delineated layer structures in human DLPFC and adult mouse brain tissues. STCGCar is a powerful tool for spatial domain identification, showcasing its effectiveness and scalability on diverse datasets. It is freely available at https://github.com/plhhnu/STCGCar.