Structural-information guided fusion for spatial domain identification from spatial transcriptomics

Accurate spatial domain identification is essential for understanding tissue organization and pathological mechanisms in spatial transcriptomics. However, existing methods mainly rely on expression profiles and spatial coordinates. Intercellular interactions are often overlooked. At the same time, preserving both local neighborhood continuity and global topological structure remains difficult.

We propose SGFST (Structural-information Guided Fusion for spatial domain identification from Spatial Transcriptomics), a novel framework for spatial domain identification in spatial transcriptomics. SGFST integrates a spatial graph and a signal graph, and employs a dual-branch graph convolutional network with attention-based fusion to capture complementary spatial and functional information. In addition, SGFST jointly optimizes a Bayesian personalized ranking loss, a zero-inflated negative binomial loss, and a distance structural information constraint to preserve local neighborhood continuity, reconstruct expression signals, and maintain global topological consistency. Experimental results on multiple datasets demonstrate that SGFST outperforms several state-of-the-art methods in spatial domain identification.

The code of SGFST is available at Github (https://github.com/xkmaxidian/SGFST) and Zenodo (DOI: 10.5281/zenodo.20624899).