The problem. Detecting spatial domains — coherent tissue regions like cortical layers or tumor niches — needs more than expression clustering, which ignores where cells are. Neighbouring spots should tend to share a domain, and the histology image carries structure too. How do you fuse expression, physical proximity, and tissue morphology into a single, spatially-aware clustering?

The idea. SpaGCN builds a graph over spots weighted by both spatial distance and histological similarity (from the H&E image), then runs a graph convolutional network that smooths each spot’s expression over its spatial-histological neighbourhood before clustering. The result respects tissue geometry — domains come out spatially contiguous — and the method also identifies the spatially variable genes marking each domain.

Why it matters. This is the graph-neural-network arriving in spatial analysis, and it ties threads together: graph convolutions descend from the deep-learning arc (AlexNet → the broader neural-net toolkit), applied to the spatial-domain problem BANKSY (day 4) also tackles. Benchmarked on the Maynard DLPFC data (earlier today), it’s a canonical spatial-domain method — exactly the kind the STU would evaluate for recovering known tissue architecture.

Verdict. Foundational for deep-learning-based spatial domain detection; fusing histology is its distinctive move. As with any GNN, results depend on graph construction and parameters. Read it as the image-plus-expression, graph-based answer to “what are the regions in this tissue.”