Fig. 2

Performance of scAI and its comparison with MOFA using eight simulated datasets. a 2D visualization of cells by applying UMAP to scRNA-seq, scATAC-seq, and aggregated scATAC-seq data obtained from scAI. Each row shows one example of each scenario from the simulated datasets. Cells are colored based on their true labels. b Cells are visualized by VscAI. c Accuracy of scAI (evaluated by AUC) in reconstructing cell loading (blue color), gene loading (orange color), and locus loading (yellow color) matrices, respectively. For each scenario, we generated a set of simulated data using five different parameters, which are indicated on the x-labels. The numbers outside and inside the brackets represent the parameters in the simulated scRNA-seq and scATAC-seq data, respectively. We applied scAI to each dataset 10 times with different seeds and then calculated the average AUCs with respect to the ground truth of the loading matrices. Datasets 5 and 6 were generated based on real datasets, which do not have ground truth of the gene/locus loading matrices. d Variance explained by each latent factor (LF) using MOFA. e Comparison of the accuracy (evaluated by normalized mutual information, NMI) of scAI and MOFA in identifying cell clusters