Separable Spatial Single-cell Transcriptome Representation Learning via Graph Transformer and Hyperspherical Prototype Clustering

Spatial transcriptomics enables the exploration of cell types, spatial domain organization, and cell–cell communication at tissue scale, serving as a powerful link between tissue morphology and molecular function. However, existing methods still struggle with limited spatial resolution, poor robustness in multi-slice alignment, and suboptimal ligand-receptor (L-R) detection in complex tissues. To address these challenges, we propose S3RL (Separable Spatial Single-cell transcriptome Representation Learning via Graph Transformer and Hyperspherical Prototype Clustering), a unified framework that integrates gene expression, spatial coordinates, and histological image features via a graph neural network and hyperspherical prototype-based separable representation learning. In spatial clustering tasks, S3RL improves the average Adjusted Rand Index (ARI) by nearly 120% on the Nanostring lung cancer dataset and over 26% on the 10X DLPFC dataset compared to state-of-the-art methods. For multi-slice spatial alignment, S3RL achieves an average ARI improvement of over 65.4% in partial brain slices and over 48.3% in complete slices relative to GraphST, highlighting its superior robustness and alignment consistency across heterogeneous spatial inputs. Moreover, S3RL reveals more biologically meaningful ligand-receptor signaling relationships, enhancing the interpretability of cell-cell communication patterns. Together, these results demonstrate S3RL’s effectiveness in enhancing spatial resolution, cell type identification, and biological insight across diverse spatial transcriptomics datasets.