Graph Contrastive Learning for Deciphering Spatial Heterogeneity in Breast Cancer

Abstract

Breast cancer exhibits significant spatial and molecular heterogeneity. Traditional bulk and single-cell transcriptomic analyses struggle to fully reveal the molecular states and microenvironmental differences across distinct tumor regions due to their inability to preserve spatial tissue information. This study proposes and applies stCL, a spatial transcriptomics domain identification framework based on graph attention networks and contrastive learning, to dissect spatial transcriptomic heterogeneity in breast cancer. It integrates gene expression and local spatial topology through a graph attention-based multi-view encoder, jointly optimizing contrastive learning loss, spatial regularization loss, and zero-inflated negative binomial reconstruction loss to learn biologically meaningful low-dimensional embeddings. Results show that stCL effectively identifies spatially coherent domains such as Tumor, Invasive, Surrounding Tumor, and Healthy regions, which largely align with pathological annotations. Quantitative evaluation indicates that stCL achieves an adjusted Rand index (ARI) of 0.6008 and normalized mutual information (NMI) of 0.7095, outperforming several existing spatial clustering methods. Further differential expression and functional enrichment analyses reveal distinct molecular signatures and functional differences among spatial domains: the Invasive region is enriched for epithelial tumor and cell proliferation-related pathways, the Surrounding Tumor region shows enrichment in immune response and extracellular matrix-related signals, while the Tumor region displays specific metabolic and epithelial characteristics.