Ó³»­´«Ã½

Tissue phenotypes, such as metabolic states, inflammation, and tumor properties, emerge from both molecular states and spatial cell organization. Spatial molecular assays provide an unbiased view of tissue architecture, enabling phenotype prediction. Graph neural networks (GNNs) offer a natural framework for analyzing spatial proteomics by integrating expression profiles with structure. We apply GNNs to classify tissue phenotypes using spatial cell patterns. We show that for relatively simple classification tasks, such as tumor grading in breast cancer, incorporating spatial context does not significantly improve predictive performance over models trained on single-cell or pseudobulk representations. However, GNNs capture meaningful spatial features, retaining prognostic signals beyond tumor labels, highlighting tumor-grade-specific cell type interactions, and uncovering complex immune infiltration patterns in colorectal cancer not detectable with traditional approaches. These findings suggest that while spatial dependencies may not always enhance classification performance in small datasets, GNNs remain valuable tools for characterizing tissue organization and interactions.