CoCo-ST detects global and local biological structures in spatial transcriptomics datasets.

Abstract

Spatial domain detection methods often focus on high-variance structures, such as tumour-adjacent regions with sharp gene expression changes, while missing low-variance structures with subtle gene expression shifts, like those between adjacent normal and early adenoma regions. Here, to address this, we introduce 'compare and contrast spatial transcriptomics' (CoCo-ST), a graph contrastive feature representation framework. By comparing a target sample with a background sample, CoCo-ST detects both high-variance, broadly shared structures and low-variance, tissue-specific features. It offers technical advantages, including multisample integration, batch-effect correction and scalability across technologies from spot-level Visium data to single-cell Xenium Prime 5K and subcellular Visium HD data. We benchmarked CoCo-ST against ten state-of-the-art spatial-domain-detection algorithms using mouse lung precancerous samples, demonstrating its superior ability to identify low-variance spatial structures overlooked by other methods. CoCo-ST also effectively distinguishes cell clusters and niche structures in Visium HD and Xenium Prime 5K data. CoCo-ST is accessible at GitHub ( https://github.com/WuLabMDA/CoCo-ST ).