Fig. 5

Sfaira allows streamlined embedding models training across tissues and on whole atlases. a, b Pre-trained embedding models can perform meaningful reconstruction of cells in held out data sets. UMAP based on latent space of the best embedding model data for pancreas data from humans (a) and mice (b). The superimposed colors correspond to the original, non-streamlined, cell type annotation. c, d Reconstruction performance comparison of different embedding models across organs and organisms. The negative binomial likelihood is used as a reconstruction performance metric on reconstructed test data of held-out test-data sets from PCA, linear, non-negative matrix factorization (nmf), autoencoder (ae), and variational autoencoder (vae) models on human (c) and mouse (d) organs. e, f Sfaira allows for training of embedding models using very large data sets. UMAP of the latent space of an embedding model trained on all mouse data in the sfaira data zoo with the data set (e) and cell type (f) superimposed