Combining Spatial Multi‐Omics Data to Decipher Spatial Domains and Elucidate Cell Heterogeneity Based on Self‐Supervised Graph Learning

Spatial multi-omics technologies enable in situ molecular profiling but face challenges in integrating multi-modal data for spatial domain identification and cell heterogeneity analysis. We propose SOTMGF, a self-supervised, goal-directed multi-view graph fusion framework for spatial multi-omics data. SOTMGF includes five modules: pre-clustering, sparse feature processing, multi-view feature extraction and fusion (integrating molecular expression, spatial location, disease microenvironment, and molecular associations), and multi-modality integration. The self-training process and graph embedding are optimized iteratively within a unified framework, enabling mutual benefits across several components. SOTMGF outperformed existing methods in spatial domain identification, data denoising, and detection of spatially variable molecular features. Innovatively, it jointly analyzes spatial transcriptomics (ST) and proteomics (SP) from the same tissue, computationally generates spatial ATAC-seq via Tangram, reconstructs spatial pseudo-expression to identify spatial dark genes/proteins (SDGs/SDPs), and iteratively optimizes self-training and graph embedding in a unified framework. SOTMGF outperforms existing methods in spatial domain detection and denoising, reveals mRNA-protein discordance, predicts key transcription factors, and aids biomarker and therapeutic target discovery, advancing spatial biology research, molecular regulatory mechanisms, and therapeutic discovery.