SPACE-GM: Decoding Disease-Relevant Microenvironments Using Geometric Deep-Learning

Tumor microenvironments are complex and heterogeneous, containing cancer cells, non-cancer cells, and a variety of immune cell types. Previous work on tumor microenvironments assigned cells to neighborhoods according to the cell-type compositions of their intermediate cellular neighbors. However, such an approach can overlook important local spatial relationships between specific cell types. Significant strides have been made in cellular property modeling by employing graph neural networks for cell-type prediction and tissue structure detection, but these techniques have not addressed how to use rich spatial transcriptomics and proteomics data to identify diseaserelevant microenvironments that may be useful for predicting clinical outcomes. Enable Medicine and collaborators developed the SPACE-GM geometric deep learning framework to address this gap.