Extended Data Fig. 3: Model ablation experiments.

Ablation experiments of TemporalVAE (without the Time-predictor) on Acinar (A, B, C), Embryo beta (D, E, F), and Human female germline datasets (G, H, I). The Ablated-TemporalVAE has the same model structure as TemporalVAE but without the Time-Predictor. The Ablated-TemporalVAE uses the same training parameters as TemporalVAE to train the ablated model on the training data to obtain their low-dimensional representations, then a linear regression (LR) model is fitted on these representations and applied to predict the pseudo-time of test data. The Spearman results (A, D, G) highlight the role of the Time-Predictor in learning temporally meaningful latent features. The distributions of the predicted time by Ablated-TemporalVAE (B, E, H) and TemporalVAE (C, F, I) on three datasets demonstrate the Time-Predictor module is essential for pseudo-time prediction.