Fig. 3

scVAEDer can be used to understand cellular dedifferentiation. a Mapping HSPC and monocyte cells into the latent prior of DDM. b Using DDM (1000 diffusion steps) and performing latent linear interpolation with 2000 equidistant samples (red dots). The absence of any sample generated in the empty region suggests that the model has learned an accurate embedding. c Heatmap showing the similarity between gene expression of generated states and the real average expression of HSPCs and monocytes (using 100 DDM steps before interpolation). d Expression of selected marker genes along the interpolation path