Spatial transcriptomics (ST) has transformed gene expression studies by integrating spatial context, offering insights from multicellular regions to subcellular levels. A key challenge in ST data analysis is the unsupervised clustering of spatial domains, crucial for identifying coherent regions and downstream analyses. Furthermore, current methods are limited to analysing single tissue slices and are ineffective for multiple vertical and horizontal ST slices, where batch effect removal is critical for integrating serial sections. Recent approaches often rely on predefined graph structures with fixed neighbourhoods, which may overlook the complex, context-dependent nature of cellular interactions. Additionally, many methods use contrastive learning, sensitive to corrupted structures and resource-intensive due to negative sampling.
To address these challenges, we developed SemanticST, a novel spatially-informed multi-semantic graph neural network. SemanticST constructs multiple semantic graphs that capture distinct contextual relationships within tissue through an unsupervised approach, overcoming fixed graph structure limitations. Additionally, SemanticST employs a community-based loss function, eliminating contrastive learning and enabling efficient, robust node representations.
We evaluated SemanticST across diverse tissue types, including human, mouse brains and breast cancer, using platforms such as Visium, Stereo-seq, and Slide-seq. SemanticST significantly outperformed recent methods in clustering accuracy, trajectory inference, and uncovering critical biological insights like tissue architecture and tumour heterogeneity. Notably, SemanticST achieved ≥10% improvement in Adjusted Rand Index and Normalised Mutual Information and was the only method to accurately identify all spatial regions and organs annotated in ground truths. Moreover, the latent representation from tissue slices of the human brain and mouse breast cancer shows that SemanticST effectively utilises spatial information to identify spatial domains across tissues while comprehensively addressing batch effects without relying on predefined batch factors. These findings position SemanticST as a powerful tool for advancing clinical and biological research, offering new avenues for precision medicine and a deeper understanding of complex tissue architecture.