Fig. 3

Comparison of clustering performance of different algorithms at various dropout probabilities and DEG numbers. a Cell-clustering by JOINT, Saver, and scImpute on a simulated dataset with three clusters (dropout probability is set to 0.3 and DEG number set to 150). Original data with no dropout is shown on the left. Adjusted Rand Index for each algorithm is shown. K-means clustering method is used for published imputation algorithms. Imputation algorithm in JOINT is used for data visualization. For datasets with dropout, we applied the PCA from the original dataset without dropout to get the 2-dimensional plot. b - c Cell-clustering scores are compared for JOINT, Saver, and scImpute algorithms at different dropout probabilities on a dataset with 150 DEG (b) and 50 DEG (c). d - e Correlation coefficients of cell-clustering results from JOINT, Saver, and scImpute to original “true labels” are averaged across all genes (Gene Correlation) or cells (Cell Correlation) at different dropout probabilities. Correlation coefficients generated from a dataset with 150 DEG (d) and 50 DEG (e) are shown