xSiGra: Explainable model for single-cell spatial data elucidation.

Area of Science:

• Computational Biology
• Bioinformatics
• Single-cell Spatial Omics

Background:

• Spatial imaging technologies enable high-resolution multi-channel images, gene expressions, and spatial locations at the single-cell level.
• Understanding cell types and their interactions within complex tissues is crucial for biological discovery.

Approach:

• Introduced xSiGra, an interpretable graph-based AI model leveraging multi-modal features from spatial imaging.
• Constructed a spatial cellular graph using immunohistology images and gene expression as node attributes.
• Employed hybrid graph transformer models and a novel Grad-CAM variant for cell type delineation and interpretable feature extraction.

Key Points:

• xSiGra effectively delineates spatial cell types and uncovers pivotal genes and cells.
• Demonstrated superior performance compared to existing methods across diverse spatial imaging datasets.
• Revealed that cellular activity is influenced by neighboring cells, not just intrinsic factors.

Conclusions:

• xSiGra provides deeper biological insights into spatial omics data.
• Identified endothelial cell subsets interacting with tumor cells, highlighting complex cellular communications.
• Facilitates understanding of heterogeneous mechanisms within tumor microenvironments.