Adv Sci (Weinh). 2026 Jun 18:e76142. doi: 10.1002/advs.76142. Online ahead of print.
ABSTRACT
Spatial transcriptomics (ST) enables the study of tissue architecture by resolving gene expression in space, but current ST platforms are constrained by limited sequencing depth and indirect single-cell identification. Existing deconvolution methods that integrate single-cell RNA sequencing (scRNA-seq) data with ST often overlook the biological principle that cells in communication with each other tend to be closer spatially. Here we introduce SPADE, a deep learning framework that aligns scRNA-seq data to spatial locations by jointly modeling expression similarity between scRNA-seq and ST data and concordance between the spot distance and cell-cell communication (CCC) patterns. SPADE also enables quantitative characterization of CCC across spots and regions. Evaluations on 55 simulated and real datasets show that SPADE achieves strong performance in recovering region-specific cell-type patterns and enhancing spatial gene expression profiles compared with existing methods. In the breast cancer datasets, SPADE demonstrates a unique advantage in identifying tumor-infiltrating immune cells and tertiary lymphoid structures. In the colorectal cancer liver metastasis dataset, SPADE distinguishes tumor heterogeneity with region-specific CCC events and describes the general CCC landscape in the tissue. Overall, SPADE highlights the key role of spatially constrained CCC in shaping tissue organization and enables biological interpretation of spatial transcriptomics data.
PMID:42314058 | DOI:10.1002/advs.76142
