Hypergraph-driven spatial multimodal fusion for precise domain delineation and tumor microenvironment decoding

Recent advancements in spatial transcriptomics have transformed tumor microenvironment research by providing insights into cellular interactions and spatial heterogeneity. A fundamental challenge is the precise delineation of spatial domains. However, existing methods remain limited in accurately identifying spatial domains, partially due to their reliance on single-view features. Moreover, these methods often struggle with many-to-many spot relationships, such as shared biological functions. To this end, we propose HAST, a hypergraph-driven spatial multimodal fusion tool for precise domain delineation and tumor microenvironment decoding. HAST integrates gene expression, spatial coordinates, and histological features to construct local hypergraphs that effectively model many-to-many spatial relationships. These local hypergraphs are dynamically aggregated into a global hypergraph, capturing higher-order interactions. To learn discriminative and biologically meaningful representations, we employ a hypergraph convolutional network, coupled with self-supervised contrastive learning, to fuse multi-view information. Extensive benchmarking across multiple datasets demonstrates that HAST outperforms state-of-the-art methods, accurately delineating spatial domains and uncovering domain-associated genes. Functional enrichment analyses further reveal biologically relevant pathways and provide novel insights into tumor microenvironment. In summary, HAST is a robust framework for decoding the spatial complexity of tumors, paving the way for precise spatial omics analyses in cancer research.