Skip to main content
Fig. 3 | Genome Biology

Fig. 3

From: scTour: a deep learning architecture for robust inference and accurate prediction of cellular dynamics

Fig. 3

scTour reconstructs the cellular dynamics of unseen cellular states in pancreatic endocrinogenesis. a UMAP visualizations of the latent space from scTour based on 3696 cells from [23], colored by cell types (top), and pseudotime (bottom). b Top: UMAP representation showing the pseudotime from the model trained with the Fev + cells excluded. Bottom: scatter plot comparing the pseudotime estimates (y-axis) with those inferred from the full dataset (x-axis), with Spearman’s ρ and p-value between them shown. c Top: UMAP representation displaying the predicted pseudotime for the held-out Fev + cells (dotted circle). Bottom: scatter plot comparing the prediction (y-axis) with the ground truth (x-axis) highlighted in orange, with Spearman’s ρ and p-value between pseudotime in x and y axes shown. Bottom inset: ROC curve for the binary classifications of endocrine progenitors versus Fev + cells (green) and Fev + versus terminal cells (blue) based on the predicted (Fev + cells) and inferred (remaining cells) pseudotime. d Streamline visualizations of the vector field from the Fev + -excluded model (top), and the prediction for these held-out cells (bottom). e Schematic depicting the model training with the Ngn3high EPs excluded, followed by prediction of their latent space given their expected developmental time. f UMAP visualization based on the reconstructed latent representations for the held-out Ngn3high EPs (red outline) and those inferred from training cells, colored by pseudotime and cell identities (inset). g As with f, but with true Ngn3high EPs (blue outline) incorporated. h Box plot displaying the Euclidean distances between the reconstructed latent representations for Ngn3high EPs and those from each cellular state, with the medians, interquantile ranges, 5th, 95th percentiles indicated by center lines, hinges, and whiskers, respectively. i Unsupervised hierarchical clustering of the reconstructed Ngn3high EPs and all the other cells based on their Euclidean distances in the latent space. Colors mark the cell types, and the reconstructed (red), true (blue) Ngn3high EPs, and remaining cells (grey)

Back to article page