Fig. 2: Performance evaluation on data reconstruction, data efficiency, clustering, and batch effect removal.

Performance evaluations were done on the three data sets marrow, gastrulation, and brain, and three different random seeds. We compared to MultiVI⁸, Cobolt¹¹, and scMM²⁰ where applicable. See the legend under D for color decoding and metric evaluation (arrows in plot titles). All values are presented as mean values +/- SEM. Individual data points for N = 3 are plotted as black dots. A, F Lower is better. A Reconstruction performance on the test RNA data measured by RMSE. B–D, G Higher is better. B Comparison of the reconstruction performance on the test set ATAC data as the AUPRC of binarized data. C Clustering performance of the train representation as the ARI based on clustering derived from the GMM for multiDGD and Leiden clustering for MultiVI. The Leiden algorithm is adjusted for the number of clusters (see Methods). D Batch effect removal of marrow and gastrulation data calculated as 1 − ASW. Brain data annotation contained no batch information. E Data efficiency was evaluated by training bone marrow models on a range of subsets. Test loss ratios were computed for models trained on three random seeds (N = 3). F, G Feature efficiency on the mouse gastrulation test set (N = 5686 cells) was investigated by training multiDGD and MultiVI on the mouse gastrulation data with (in 5%) and without feature selection (all). Performance values were only evaluated on the smaller feature set for comparability. Asterisks indicate significance based on two-sided Mann-Whitney U (MWU) tests (N = 5686). All values are provided in the Source Data. F RMSE for RNA reconstruction performance. MWU test in 5% (p-value 1e-202), all (p-value 0.000). G AUPRC for ATAC reconstruction performance. MWU test in 5% (p-value 0.062), all (p-value 0.021).