MLN2SVG: domain-aware spatially variable gene detection using contrastive variational autoencoder and multi-level...

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatial transcriptomics (ST) enables gene expression analysis in intact tissues.
• Identifying spatially variable genes (SVGs) is challenging due to data sparsity and heterogeneity.
• Existing methods struggle with domain-level dependencies in spatial data.

Purpose of the Study:

• To develop a domain-aware framework for joint discovery of tissue domains and SVGs.
• To improve the accuracy and biological interpretability of SVG detection in ST data.
• To address limitations of existing methods in handling spatial heterogeneity and sparsity.

Main Methods:

• Proposed MLN2SVG, integrating contrastive variational autoencoding with a multi-level neighbor (MLN) search algorithm.
• Constructed a weighted spatial graph to capture local and long-range spatial relationships.
• Employed a deep contrastive variational autoencoder for data representation alignment and biological diversity preservation.

Main Results:

• MLN2SVG outperformed existing methods in clustering accuracy, robustness, and biological interpretability across human and mouse ST datasets.
• Successfully identified fine-grained spatial organization in breast cancer tissues, including tertiary lymphoid structures.
• Delineated region-specific immune architectures within different compartments of breast cancer tissues.

Conclusions:

• MLN2SVG provides a robust and biologically interpretable framework for spatial transcriptomics data analysis.
• The integration of spatial domain discovery and SVG detection enhances the understanding of tissue complexity.
• MLN2SVG offers a powerful tool for uncovering molecular and structural organization in tissues.