xSiGra: explainable model for single-cell spatial data elucidation

Area of Science:

• Computational Biology
• Bioinformatics
• Artificial Intelligence in Medicine

Background:

• Spatial imaging technologies provide high-resolution single-cell data, including gene expression and spatial locations.
• Understanding complex cellular interactions and identifying distinct cell types is crucial for biological insights.

Purpose of the Study:

• To introduce xSiGra, an interpretable graph-based AI model for elucidating spatial cell types and their features.
• To leverage multimodal data from spatial imaging for enhanced biological discovery.

Main Methods:

• Constructing a spatial cellular graph using immunohistology images and gene expression data.
• Employing hybrid graph transformer models for spatial cell type delineation.
• Integrating a gradient-weighted class activation mapping variant for interpretable feature identification.

Main Results:

• xSiGra demonstrates superior performance compared to existing methods across diverse spatial imaging datasets.
• Analysis of a lung tumor slice revealed cell importance scores, highlighting the impact of neighboring cells on cellular activity.
• Identified interpretable genes uncovered interactions between endothelial cells and tumor cells, suggesting complex mechanisms.

Conclusions:

• xSiGra provides a powerful tool for analyzing complex spatial omics data.
• The model facilitates deeper biological insights by identifying pivotal genes and cell-cell interactions.
• This approach enhances our understanding of cellular heterogeneity and microenvironment dynamics.