Fig. 5

Smoother enables spatially aware joint embeddings of single-cell and Slide-seqV2 data of human prostate and improves reference mapping accuracy. a Schematic overview of the spatial conversion of non-spatial auto-encoder models in the Smoother framework. Smoother enforces spatial consistency via the detachable loss function. This allows the same model to be trained and applied to both spatial and non-spatial data, generating a joint spatially aware embedding. b UMAP visualization of the latent representation of scRNA-seq data of human prostate from the Tabula Sapiens [48], colored by tissue compartment (left) and technical batch (right). The representation was generated from a pretrained RNA-only SCVI prostate model [49]. c Violin plots showing the number of expressed gene (left) and total RNA counts (right) per cell or spot in data of different technologies. d UMAP visualization of the joint latent representation of the Tabula Sapiens prostate scRNA-seq reference and the Slide-seqV2 data of a healthy prostate section [50]. Following the SCVI data integration workflow, the RNA-only model was fine-tuned on the query spatial data with unfrozen parameters to mitigate batch effect. e Spatial visualizations of the tissue compartment (left) and cell type prediction (right) results based on the joint RNA-only embeddings shown in d. f UMAP visualization of the joint latent representation generated by SpatialVAE. The spatially aware model has the same architecture as RNA-only models in b and d, except it was fine-tuned to minimize the proposed spatial loss in addition to the original reconstruction and KL losses. g Spatial visualizations of the tissue compartment (left) and cell-type prediction (right) results based on the joint RNA-only embeddings shown in f