Fig. 5

Evaluation of JOINT’s performance in DEG analysis. a - d Comparison of the performance of DEG analysis algorithms when cell labels are known and different dropout probabilities are assigned to each cell-cluster. AUC scores for MAST, scDD, DESeq2, and JOINT when different dropout probabilities are assigned to each cell-cluster in datasets with 50 DEG (a), 100 DEG (b) and 150 DEG (c) are shown. d ROC curves for MAST, scDD, DESeq2, and JOINT when mean dropout probability for all cells is set to 0.1 (dropout probability varies by 0.05 for each cell-cluster) and DEG number is set to 150. e - h Comparison of the performance of different DEG analysis algorithms when cell labels are unknown and the same dropout probability is assigned to all cells. AUC scores for MAST, scDD, DESeq2, and JOINT when the dropout probability is set to the same value for all cells in datasets with 50 DEG (e), 100 DEG (f) and 150 DEG (g) are shown. h ROC curves for MAST, scDD, DESeq2, and JOINT when mean dropout probability for all cells is set to 0.1 and DEG number is set to 150. i AUC curves of DEG analysis algorithms in combination with imputation methods and JOINT are shown. j Computing time of one iteration of the JOINT EM algorithm when run by TensorFlow using GPU, TensorFlow using CPU (run on compiled C code), and Python-based NumPy implementation using CPU. Computing time is tested for different numbers of genes