STELLA: a spatial transcriptomics framework for microenvironment decoding using dynamic graph neural networks

Spatial transcriptomics technology can analyze gene expression while retaining spatial information, but it is still challenging to accurately identify spatial domains and decode intercellular communication networks. This study proposes the STELLA framework, which integrates dynamic graph neural networks and self-supervised learning strategies to analyze spatial transcriptome data. STELLA constructs a complementary space-feature dual-graph structure, optimizes connection weights through dynamic adjacency matrix learning, and independently encodes spatial and expression information through a dual-channel graph convolutional network. The multi-head attention mechanism adaptively integrates different information sources, and feature permutation contrast learning improves representation discrimination capabilities. In a systematic evaluation across multi-platform datasets, STELLA accurately identified the tumor-muscle interface region in zebrafish melanoma, suggesting a potential association between the MT-CO1-mediated mitochondrial electron transport pathway and invasion front processes. It also identified immune aggregation areas in intestinal tissues and implicated the cyclosporin A signaling pathway in the FDCSP⁺ S4 stromal cell microenvironment. Additionally, STELLA detected that PERIOSTIN and MHC-II form a central-peripheral bidirectional regulatory network with complementary directionality in the mouse striatum. Finally, it analyzed the transition from WSN-mediated early stress response to reaEGC-mediated tissue reconstruction during axolotl brain regeneration.