A Graph Convolutional Network for Spatial Transcriptomics Modelling

Single-cell spatial transcriptomics measures gene expression for individual cells and the positions of these cells within a tissue sample. This data provides a new lens for studying cell-cell communication, but how to most efficiently use this data for inferential purposes is still being investigated. How cells interact with their neighbors can be investigated by comparing models where cells are aware of their neighbors against models where cells are not aware of their neighbors. Posing paired models offers one way to interpret spatial transcriptomics data, but this approach can fail when prediction algorithms are insufficiently flexible, either due to a reliance on fixed-dimensional neighboring expression encodings as model inputs or due to the limited expressivity of the prediction rule that maps neighborhood encodings to predictions. To obtain a sufficiently flexible model, we developed DeepST, a graph convolutional network that learns on graphs defined from spatial transcriptomics data sets. The contrast between DeepST’s predictions and the predictions from a baseline regressor lacking access to cell neighborhood information provides insights into how cells interact.