Spatial omics technologies are revolutionizing our understanding of cellular organization and function. The need to accurately analyse the resulting data has motivated the development of new methods for the essential tasks of identifying cell types1 and tissue domains2. However, which approach works best for either task is unclear; hence, many spatial omics studies still rely on traditional, non-spatial clustering tools to identify cell types. Moreover, few methods directly address the important problem of spatially aware batch correction across samples or datasets. Additionally, most current algorithms struggle to scale beyond 100,000 cells, which limits their application to the latest spatial technologies that can profile over a million cells in a single dataset. Thus, there is a need for algorithms that combine spatial locations and molecular abundance for labelling cells without compromising scalability.

The spatial clustering strategy used by BANKSY was validated through its ability to detect niche-dependent cell states with subtle gene expression differences, which were corroborated with single-cell RNA sequencing data. For instance, BANKSY stratified mature oligodendrocytes in the mouse hypothalamus into white matter (anterior commissure) and grey matter (general hypothalamic area) subpopulations. In a mouse hippocampal area dataset, it distinguished subpopulations of neurons in the somatosensory cortex and CA3 regions that could not be identified using other methods.