Spatially Varying Graphical Models for Cell-Cell Interaction Networks in Multiplexed Tissue Imaging

Area of Science:

• Computational Biology
• Immunology
• Bioinformatics

Background:

• Multiplexed imaging provides single-cell resolution of tumor immune microenvironments.
• Current methods for analyzing cell interactions lack spatial context and conditional analysis.
• Understanding cell-cell networks is vital for characterizing complex biological systems.

Purpose of the Study:

• To develop a Bayesian framework (GP-GHS) for inferring spatially varying cell-cell interaction networks from multiplexed imaging data.
• To accurately model spatial heterogeneity in cell interactions within tissue compartments.
• To identify differential cell-cell interactions associated with specific disease subtypes.

Main Methods:

• Utilized a Bayesian nodewise regression framework with Gaussian Processes (GP) over the tissue domain.
• Employed a Hilbert Space Gaussian Process (HSGP) expansion for computational scalability.
• Incorporated a group horseshoe prior for robust edge inference and spatial smoothness.
• Implemented a block Gibbs sampler for efficient posterior inference.

Main Results:

• GP-GHS significantly outperformed existing methods in simulation studies across various metrics (F1, MCC).
• The group shrinkage prior was identified as a critical component for accurate network recovery.
• Analysis of colorectal cancer data revealed 13 differentially active cell-cell interactions.
• Identified a Treg-centered immunosuppressive network, particularly active in the diffuse inflammatory subtype.

Conclusions:

• GP-GHS provides a powerful and scalable method for inferring spatially structured cell-cell interaction networks.
• The identified Treg-centered network offers insights into colorectal cancer immunosuppression mechanisms.
• This approach enhances the characterization of tumor immune microenvironments and disease subtypes.