Fig. 3

Imputation accuracy in the down-sampling experiment. Results are shown for down-sampling experiments using either expression-dependent dropout rate (a) or expression-independent dropout rate (b) for four different data sets (Gurn, Cell Type, Time Course, and Shalek). Imputation accuracy are measured by comparing imputed values to the original truth and are evaluated for two different types of zeros separately: zeros that are due to low expression level in the original data and the multinomial subsampling step (down-sampling entries; y-axis) and zeros that are due to dropout events (dropout entries; x-axis). Accuracy is measured by correlation for the dropout entries and by L1 loss for the down-sampling entries (because of an excess of zero values in the original data for the down-sampling entries). Color of the dots represents methods for comparison: DrImpute (blue), MAGIC (green), SAVER (pink), scImpute (purple), VIPER with elastic net selection (peach), and VIPER with lasso selection (dark blue). Shape of the dots represents the down-sampling rate used in the multinomial subsampling step