stGCL: a versatile cross-modality fusion method based on multi-modal graph contrastive learning for spatial transcriptomics

TL;DR

The paper introduces stGCL, a new method that combines different types of data from spatial transcriptomics to better understand tissue structure and function.

Contribution

stGCL introduces a novel cross-modality fusion framework using graph contrastive learning for spatial transcriptomics data integration.

Findings

stGCL reliably identifies spatial domains in tissue data.

The method supports integration of vertical and horizontal tissue slices.

stGCL is generalizable across different platforms and resolutions.

Abstract

Advances in spatial transcriptomics have enabled high-resolution mapping of tissue architecture at the molecular level, yet integrating its multi-modal data remains challenging. Here, we present stGCL, a framework for accurate and robust integration of gene expression, spatial coordinates, and histological features. stGCL employs a histology-based Vision Transformer to extract morphological features and a multi-modal graph autoencoder with contrastive learning for cross-modal fusion. In addition, we introduce a spatial coordinate correction and registration strategy to support multi-slice integration. We demonstrate that stGCL reliably identifies spatial domains, integrates vertical and horizontal tissue slices, and highlight its generalizability across platforms and resolutions.

The online version contains supplementary material available at 10.1186/s13059-025-03896-w.

Figures