Fig. 1

Overview of DISC. a DISC contains an autoencoder, a recursive predictor, an imputer to compute an imputation expression profile, and a reconstructor to compute a reconstructed expression profile. b DISC is trained in a semi-supervised manner: (1) the imputer learns the expression of positive-count genes, (2) the reconstructor learns both the expression of positive-count genes and the pseudo expression of zero-count genes assigned by the imputer, and (3) the predictors learn both the expression of positive-count genes and the pseudo expression of zero-count genes assigned by the decoder of the same step. c Compression module reduces the large latent representations of multiple steps into a much smaller dimension for visualization and clustering. d T-distributed Stochastic Neighbor Embedding (T-SNE) visualization and clustering using top 30 PCs generated by PCA transformation from the selected top 2000 highly variable genes (HVGs) of the RETINA dataset (ACC = 0.950). e T-SNE visualization and clustering using 50 latent features generated by the compressor of DISC from all 14,871 genes (without HVGs selection) of the RETINA dataset (ACC = 0.944)