STGAT: spatial domain identification of consecutive slices based on graph contrastive learning

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomic technologies generate multi-tissue datasets, necessitating advanced computational integration methods.
• Existing methods for spatial domain identification struggle to fully utilize spatial context and gene expression across sections.
• Current graph contrastive learning frameworks often introduce noise through positive and negative sample pair construction.

Purpose of the Study:

• To introduce STGAT, a novel framework for spatial domain identification in multi-tissue spatial transcriptomic data.
• To overcome limitations of existing methods by leveraging spatial context and gene expression across sections.
• To enable effective self-supervised representation learning using only positive sample pairs in graph contrastive learning.

Main Methods:

• Precise spatial alignment of multi-tissue sections using gene expression similarity.
• A graph contrastive learning strategy utilizing only positive pairs within a unified spatial domain.
• Self-supervised representation learning for graph nodes.

Main Results:

• STGAT significantly enhances clustering accuracy in spatial domain identification across multi-section and cross-technology datasets.
• Application to mouse olfactory bulb sections reveals sharply defined spatial domain boundaries and accurate anatomical region identification.
• The framework provides a more refined characterization of the tumor microenvironment.

Conclusions:

• STGAT offers a robust and accurate approach for spatial domain identification in complex spatial transcriptomic datasets.
• The method effectively integrates spatial context and gene expression information across multiple sections.
• STGAT advances the analysis of spatial transcriptomics, with implications for understanding tissue organization and disease.