SpaVGMC: A Unified Representation Learning Framework via Structural and Semantic Alignment in Spatial Transcriptomics

Spatial transcriptomics technologies profile gene expression within its native spatial context, offering new insights into tissue organization and disease. However, accurate spatial domain identification remains challenging due to local oversmoothing, impaired topological fidelity, and insufficient modeling of global semantic structures. To address these challenges, we propose SpaVGMC, a unified representation learning framework that jointly models structural dependencies and transcriptional semantics. The framework integrates structured variational representation learning, structural information alignment, and semantic alignment, allowing the capture of probabilistic uncertainty, multiscale spatial dependencies, and global transcriptional organization. Specifically, SpaVGMC formulates representation learning as a structured variational inference process with context-aware message-passing. A structural information alignment mechanism preserves topological fidelity by aligning latent embeddings with the spatial graph via mutual information at both the edge and neighborhood levels. In addition, a semantic alignment mechanism organizes representations according to transcriptional similarity through distribution-aware contrastive learning without requiring data augmentation. By jointly modeling representations, structures, and semantics, SpaVGMC learns robust, discriminative, and biologically interpretable embeddings. Extensive experiments across diverse spatial transcriptomics data sets demonstrate that SpaVGMC consistently outperforms state-of-the-art methods in spatial domain identification, showing improved agreement with tissue structures and enhanced detection of fine-grained subdomains. Collectively, these results establish SpaVGMC as a robust and scalable framework for spatial omics analysis.