SemanticST: A Scalable Multi-Contextual Graph Learning Framework for Uncovering Spatial Niches and Robust...

Spatial transcriptomics (ST) analysis is often hindered by technical limitations and methodological biases that let dominant signals overshadow subtle but crucial biological patterns, such as rare cell types and fine-grained heterogeneity. This is especially true for high-complexity datasets from platforms like Xenium. We present SemanticST, a graph neural network (GNN) framework that addresses this challenge through a fundamentally different design. SemanticST is the first GNN method to implement mini-batch training, enabling scalable analysis of massive ST datasets (validated on Xenium). Crucially, it employs a multi-semantic graph fusion strategy that learns disentangled biological representations across tissue, using a min-cut loss that requires neither graph corruption nor contrastive sampling. Benchmarking across diverse tissues (e.g., brain, embryo, tumor) confirms consistent superiority. It achieves up to 20% higher ARI/NMI on the gold-standard brain cortex and uniquely delineates all mouse olfactory bulb layers and hippocampal sub-regions. In high-resolution breast cancer data, SemanticST identifies computationally plausible spatial domains, including a candidate rare triple receptor-positive region and a FOXC2-enriched EMT-associated domain, from Xenium alone. Furthermore, SemanticST provides superior, robust multi-sample integration on established benchmarks. SemanticST offers an essential, scalable framework for translating spatial complexity into biologically informative and testable hypotheses.