Spatial transcriptomic data reveals pure cell types via the mosaic hypothesis

Neurons display remarkable diversity in their anatomical, molecular, and physiological properties. Although observed stereotypy in subsets of neurons is a pillar of neuroscience, clustering in high-dimensional feature spaces, such as those defined by single cell RNA-seq data, is often inconclusive and cells seemingly occupy continuous, rather than discrete, regions. In the retina, a layered structure, neurons of the same discrete type avoid spatial proximity with each other. While this principle, which is independent of clustering in feature space, has been a gold standard for retinal cell types, its applicability to the cortex has been only sparsely explored. Here, we provide evidence for such a mosaic hypothesis by developing a statistical point process analysis framework for spatial transcriptomic data. We demonstrate spatial avoidance across many excitatory and inhibitory neuronal types. Spatial avoidance disappears when cell types are merged, potentially offering a gold standard metric for evaluating the purity of putative cell types.