Identifying spatial domains of spatially resolved transcriptomics via multi-view graph convolutional networks.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Spatially resolved transcriptomics (ST) technologies provide gene expression data with spatial context.
• Understanding the tissue microenvironment requires linking gene expression to spatial distribution.
• Identifying spatial domains from ST data remains a computational challenge.

Purpose of the Study:

• To propose a novel unsupervised learning framework, STMGCN, for accurate spatial domain identification.
• To effectively integrate gene expression data with spatial information for enhanced biological insights.
• To develop a computational tool for analyzing spatial transcriptomics data.

Main Methods:

• Constructing multiple neighbor graphs (views) using different similarity measures based on spatial coordinates.
• Employing multi-view graph convolution networks (MGCNs) to learn view-specific gene expression embeddings.
• Utilizing an attention mechanism to adaptively fuse embeddings for final spot representation.

Main Results:

• STMGCN achieved competitive performance in spatial domain identification across diverse ST datasets and platforms.
• The framework successfully identified spatially variable genes with domain-enriched expression patterns.
• Demonstrated superior performance compared to five state-of-the-art spatial and non-spatial methods.

Conclusions:

• STMGCN is a powerful and efficient computational framework for spatial domain identification in ST data.
• The method effectively leverages spatial context to improve the analysis of tissue microenvironments.
• STMGCN enhances the expressive power of latent embeddings through multi-view graph convolutions.