M‐STGCN: A Position‐Aware Multimodal Graph Convolutional Framework for Joint Spatial Domain Identification and Gene Expression Denoising

Spatially transcriptomics (ST) has revolutionized our ability to profile gene expression within the architectural complexity of tissue microenvironments. However, decoding spatial heterogeneity requires robust multimodal data integration that unifies gene expression, spatial positions, and histopathological images to overcome modality‐specific biases. Here, we propose M‐STGCN, a multimodal unsupervised framework that constructs a spatial weight matrix from spatial coordinates to simultaneously refine both gene expression profiles and image features, by which we establish position‐aware feature enhancement served as a core innovation before graph fusion. Verified on human brain and breast cancer datasets, M‐STGCN significantly improves the accuracy of spatial domain identification. Ablation studies confirm the importance of its position‐aware and image modality integration. For ST platforms lacking images and at diverse resolutions, M‐STGCN maintains robust performance utilizing only gene expression and spatial coordinates. By effectively denoising raw spatial transcriptomic profiles, our approach identifies more significant spatial domain marker genes, as well as potential prognostic biomarkers for breast cancer. Moreover, M‐STGCN reveals that image features and spatial information contribute equally to breast cancer analysis, underscoring the critical role of image features. As a versatile and scalable tool, M‐STGCN enables unbiased integration of multimodal data, facilitating the deciphering of spatial heterogeneous in complex tissues.