Graph neural networks learn emergent tissue properties from spatial molecular profiles.

Area of Science:

• Computational Biology
• Bioinformatics
• Pathology

Background:

• Tissue phenotypes arise from molecular states and spatial cell organization.
• Spatial molecular assays offer insights into tissue architecture for phenotype prediction.
• Graph neural networks (GNNs) provide a framework for analyzing spatial proteomics data.

Purpose of the Study:

• To apply GNNs for classifying tissue phenotypes using spatial cell patterns.
• To evaluate the impact of spatial context on predictive performance compared to traditional methods.
• To explore the ability of GNNs to capture prognostic signals and complex biological interactions.

Main Methods:

• Utilized graph neural networks (GNNs) to integrate spatial proteomics data and cell expression profiles.
• Trained models on spatial cell patterns for tissue phenotype classification.
• Compared GNN performance with models using single-cell or pseudobulk representations.

Main Results:

• For simple tasks like breast cancer grading, spatial context did not significantly improve GNN performance over non-spatial methods.
• GNNs successfully captured meaningful spatial features and prognostic signals beyond tumor labels.
• GNNs identified tumor-grade-specific cell type interactions and complex immune infiltration patterns in colorectal cancer.

Conclusions:

• Spatial context may not always enhance classification performance in smaller datasets.
• GNNs are valuable for characterizing tissue organization and cell-cell interactions.
• GNNs uncover complex biological patterns not detectable by traditional approaches, offering new insights into tissue phenotypes.