stDyer enables spatial domain clustering with dynamic graph embedding.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Spatially resolved transcriptomics (SRT) enables gene expression analysis within tissue architecture.
• Identifying distinct spatial domains is crucial for interpreting SRT data.
• Existing methods face challenges in scalability and accuracy for complex spatial patterns.

Purpose of the Study:

• To introduce stDyer, a novel deep learning framework for accurate and scalable spatial domain clustering in SRT data.
• To enhance the interpretation of gene expression patterns within their spatial context.

Main Methods:

• stDyer integrates a Gaussian Mixture Variational AutoEncoder with graph attention networks.
• Dynamic graphs adaptively link data points based on learned embeddings and mixture assignments.
• Mini-batch processing and multi-GPU support ensure scalability for large datasets.

Main Results:

• stDyer achieved superior performance in spatial domain clustering compared to existing methods.
• The framework demonstrated effectiveness in multi-slice analysis and handling large-scale SRT datasets.
• Smoother domain boundaries and improved clustering accuracy were observed.

Conclusions:

• stDyer provides an effective and scalable solution for spatial domain identification in SRT data.
• This framework advances the analysis of gene expression in tissue microenvironments.
• stDyer facilitates deeper insights into spatial biology and disease mechanisms.