Fig. 4

Impact of imputation methods on unsupervised clustering analysis. a Heatmap of four performance metrics—entropy of cluster accuracy (H_acc), entropy of cluster purity (H_pur), adjusted Rand index (ARI), and median Silhouette index—averaged across seven datasets from CellBench [18]. Each metric shows the average performance across 7 datasets in CellBench. To compare imputation methods across metrics, the metrics were re-scaled to between 0 and 1 and the order of H_acc and H_pur were flipped to where a higher standardized score translates to better performance. Imputation methods (rows) are ranked by the average performance between the mean of the first three metrics (H_acc,H_pur and ARI) and the fourth metric (medianSil). b Dimension reduction results after applying PCA to the sc_celseq2_5cl_p1 data with no imputation (left) and with imputation using MAGIC (right). The colors are the true group labels. c Overall score (or average of the four performance metrics) for Louvain clustering (x-axis) and k-means clustering (y-axis). d–f Same as a–c except using the scRNA-seq dataset of ten sorted peripheral blood mononuclear cell (PBMC) cell types from 10x Genomics [3] (PBMC_10x_tissue dataset). White areas with black outline in d indicate that the imputation methods did not return output after 72 h. Also, e uses UMAP components [52] instead of principal components. Please refer to Additional file 1: Figures S6, S7, S9 for Louvain clustering results and metrics in each dataset and Additional file 1: Figure S8 for UMAPs of other methods