Fig. 2

Integrating snRNA-seq and snATAC-seq data from bipolar cells in a mouse retinal sample. a–c The UMAP views of the snRNA-seq profiles (a), the snATAC chromatin accessibility peak profiles (b), and the compressed gene-based chromatin accessibility profiles (c) of 1276 mouse retinal bipolar cells. The cells are colored by cell types annotated based on RNA expression levels (ground truth, Additional file 1: Fig. S3). The Adjusted Rand Index (ARI) values are labeled in each panel. Given we use RNA annotation as the gold standard, the ARI for RNA clustering in (a) is 1. d, e UMAPs generated from the bindSC-integrated snRNA and snATAC co-embeddings. Plotted respectively are cells in the snRNA-seq data (d) and those in the snATAC-seq data (e). f Consistency between cell types computationally inferred from ATAC profiles by bindSC, Seurat v3.0, LIGER, and Harmony, respectively, with the ground truth. Darkness of the dots corresponds to degree of consistency while size of the dots the fraction of cells per row. Overall accuracies are shown in the subtitle after the method names. g Pearson correlation coefficients between the imputed and the initial gene score (top panel) and the ground truth RNA profiles (bottom panel) over bindSC iterations. The result at iteration 0 corresponds to the traditional CCA method. Each dot corresponds to the accuracy of one known marker gene (The full gene list is shown in Additional file 1: Fig. S3). h Downsampling schemes to generate imbalanced datasets between snRNA-seq and snATAC-seq data. Each value in the table denotes cell number available. We generated two imbalanced data integration scenarios: (1) downsampling 50% of the five major cell types (BC1-5) while keeping all the cells from the five minor cell types (BC6-10) and (2) removing the five minor cell types (BC6-10) in the snATAC-seq data. Label transfer accuracy achieved by various methods on these imbalanced datasets are shown for scenario 1 (i) and scenario 2 (j)