An interpretable deep learning framework for revealing fine-grained spatial domains

Spatial transcriptomics enables gene expression profiling with preserved spatial context, yet accurate spatial domain delineation remains challenging. We present RIF, a robust and interpretable deep learning framework for fine-grained spatial domain identification. RIF leverages graph neural networks semi-supervised with pseudo-labels to preserve tissue structural integrity. It incorporates a dual-preprocessing ensemble strategy to improve model robustness and generalization. It further enables spatially variable gene detection through network interpretability. Across multiple sequencing platforms, RIF outperformed state-of-the-art comparison algorithms in spatial domain identification. In mouse brain tissue, RIF identified corresponding hippocampal subfields in the spatially separated dorsal and ventral regions and resolved the laminar organization of the olfactory bulb. In breast cancer, RIF-defined domains suggest suppressed T cell activity associated with hypoxic and oxidative microenvironments. In an invasive ductal carcinoma dataset of more than 720,000 cells, RIF-defined domains suggest an epithelial-mesenchymal transition gradient, including reduced hormone dependence and extracellular matrix remodeling.