Fig. 6

Interpreting blood cell types in GTEx using xCell gene sets. a Comparing BioBombe scores of all compressed latent features for variational autoencoder (VAE) models when bottleneck dimensionalities are set to k = 2 and k = 3. b Comparing mean BioBombe z-scores of aggregated latent features across two VAE models with k dimensionalities 2 and 3. Tracking the BioBombe z-scores of c “Neutrophils_HPCA_2” and d “Monocytes_FANTOM_2” gene sets across dimensionalities and algorithms. Only the top scoring feature per algorithm and dimensionality is shown. e Projecting the VAE feature k = 3 feature and the highest scoring feature (VAE k = 14) that best captures a neutrophil representation to an external dataset measuring neutrophil differentiation treatments (GSE103706). f Projecting the VAE k = 3 feature that best captures monocytes and the feature of the top scoring model (NMF k = 200) to an external dataset of isolated hematopoietic cell types (GSE24759)