Graph convolutional networks for inferring cell-cell communication from spatial transcriptomics data.

Motivation: Single-cell spatial transcriptomics provides gene expression measurements of individual cells while preserving their spatial positions within tissue. Cell-cell communication (CCC) can be inferred by comparing the predictions of held-out gene expression levels by a pair of models: one that incorporates cellular neighborhood information and another that does not. The performance gap indicates the influence of CCC. However, existing methods that adopt this general approach often rely on spatially informed models that use simplistic representations of spatial context. This reliance on such representations does not merely lead to suboptimal predictions: it undermines the validity of the model comparison itself, which hinges on the accurate estimation of conditional expectations. Results: We propose using a graph convolutional network (GCN) as a highly expressive spatially informed model, with cells as nodes and spatial proximity as edges. In semi-synthetic datasets, we show that several existing approaches relying on simplistic neighborhood features can produce spurious inferences about CCC, whereas our GCN-based approach avoids these pitfalls. In MERFISH and Xenium mouse brain tissue, our method identifies genes with known spatial variation, suggesting that it successfully infers CCC-affected genes. Availability and implementation: Code to reproduce our results is available from https://github.com/prob-ml/spice.