Fig. 4

Disentanglement and generation performance of WGAN-GP, β-TCVAE, and MichiGAN. a UMAP plots of real data colored by the ten representations of β-TCVAE and generated data colored by the ten representations of WGAN-GP and MichiGAN on the simulated data with non-linear step. The β-TCVAE panel is reproduced from Fig. 1b for clarity. b Bar plots of Spearman correlations between ten representations and each of the four ground-truth or inferred variables for WGAN, β-TCVAE and MichiGAN on the simulated data with non-linear step. The β-TCVAE panel is reproduced from Fig. 1c for clarity. c Random forest error of PCA, GMM, VAE, β-TCVAE, WGAN-GP, and MichiGAN on the whole Tabula Muris data during training. d Inception score of PCA, GMM, VAE, β-TCVAE, WGAN-GP, and MichiGAN on the whole Tabula Muris data during training. Error bars indicate standard deviation across 5 runs. For clarity, the error bars for MichiGAN are shown only for the last 100 epochs because the convergence speed in earlier epochs is variable, and the error bars for PCA and GMM are omitted because of their small and large variability