Encoding functional edges in graphs to model spatially varying relationships in the tumor microenvironment

Area of Science:

• Oncology
• Computational Biology
• Bioinformatics

Background:

• Tumor microenvironment (TME) characterization is crucial for cancer progression understanding and personalized therapy development.
• Spatial context within the TME, spanning molecular to tissue levels, is vital but challenging to model comprehensively.
• Current spatial modeling methods are often modality-specific and lack flexibility.

Purpose of the Study:

• To introduce SPIFEE, a flexible graph deep learning framework for modeling the TME across multiple biological organization levels.
• To enhance spatial insights and cross-modality integration for TME analysis.
• To improve the characterization of cellular, phenotypic, and pathway interactions within the TME.

Main Methods:

• Developed SPIFEE, a graph deep learning framework encoding spatially varying functional vectors into graph edges and representing TME entities as nodes.
• Applied SPIFEE to multiplex immunofluorescence, H&E histopathology, and spatial transcriptomics datasets.
• Integrated graph attention mechanisms to identify multi-scale spatial interactions.

Main Results:

• SPIFEE demonstrated versatility across diverse spatial omics datasets, enabling rich characterization of TME interactions.
• Function-based edge representations in SPIFEE improved performance over existing spatial modeling approaches.
• SPIFEE identified multi-scale spatial interactions significantly associated with disease state and patient survival.

Conclusions:

• SPIFEE offers a flexible and powerful graph-based approach for comprehensive TME modeling.
• The framework enhances the representational power of spatial modeling, enabling deeper interrogation of the TME.
• SPIFEE advances personalized cancer analysis by uncovering critical spatial insights relevant to patient outcomes.