stGCL: Spatial domain identification and multi-slice integration analysis in spatial transcriptomics with multi-modal graph contrastive learning

stGCL is a multi-modal graph contrastive learning framework that integrates transcriptional profiles, histological profiles and spatial information learning the spot joint embedding, enabling spatial domain detection, spatial data integration and downstream analysis. First, stGCL utilizes the pre-trained ViT model to extract the latent representation of each spot image. Then stGCL employs multi-modal graph attention auto-encoder (GATE) and contrastive learning to extract discriminative information from each modality and fuse them efficiently to generate meaningful joint embeddings. Specifically, multi-modal GATE learns spot joint structured embedding by iteratively aggregating gene expression features and histological features from adjacent spots. Furthermore, stGCL adopts contrastive learning to maximize the mutual information between each spot joint embedding and the global summary of the graph, which endows the learned joint representation with not only local but also global features. Finally, stGCL offers simple and effective vertical and horizontal integration methods for analyzing multiple tissue sections, which encourages smoothing of adjacent spot features within and between sections and mitigates the batch effect.