Spatial transcriptomics has the potential to reveal cellular interactions by measuring gene expression in situ while maintaining the tissue context of each cell.Existing deep learning methods for non-spatial single-cell omics optimize cellularembeddings of gene expression. They enable the harmonization between experimental batches while embedding the variation of the cell state. Spatial transcrip-tomics allows one to study the cell state composition of a spatial neighborhood.These cellular niches confine the tissue organization and encompass functionalunits of an organ. However, computational methods for encoding meaningful low-dimensional representations of both gene expression and cell states of neighboringcells a are currently lacking. Here, we introduce NicheVI, a deep learning modelthat decodes gene expression, niche cell-type composition, and variation in cellstate of other cells within a niche. In case studies, NicheVI uncovered additionalfine-grained heterogeneity of cell-types not captured by non-spatial and other spatially aware models and corresponding to the cellular niche