Graph Contrastive Learning for Inferring Spatial Cell Composition from Integrated Single-cell RNA Sequencing and Spatial Transcriptomics.

Spatial transcriptomics (ST) technologies integrate high-throughput transcriptomic profiling with preserved spatial cellular organization. However, many of these technologies fail to achieve single-cell resolution, hindering the identification of cell-type-specific spatial patterns and gene expression variation. We propose Graph Contrastive learning for Inferring spatial cell composition from integrated single-cell RNA sequencing (scRNA-seq) and Spatial transcriptomics (GCIRS). GCIRS constructs spot-cell heterogeneous graphs and applies metapath-based reasoning to uncover potential inter-spot connections. These inferred relationships are used to build spot-spot homogeneous graphs, effectively preserving cell-type-specific information. Furthermore, GCIRS employs a structure-aware graph contrastive learning framework to align structural patterns between real and pseudo-spot graphs, enabling cross-modality structural knowledge transfer, enhancing the representation of real spatial spots for accurate cell-type deconvolution. We evaluate GCIRS against eleven state-of-the-art methods across multiple datasets. GCIRS significantly improves spatial cell composition inference accuracy and outperforms all baselines in reconstructing cell distribution patterns. Additionally, GCIRS reveals tumor heterogeneity in human pancreatic ductal adenocarcinoma samples and delineates complex tissue structures, such as the laminar organization of the mouse cerebral cortex and the mouse cerebellum.