We next reduced the dimensionality of the DNA accessibility assay by latent semantic indexing (LSI)4,47 and reduced the dimensionality of the gene expression assay by principal component analysis (PCA). We constructed a low-dimensional visualization of the DNA accessibility assay using uniform manifold approximation and projection (UMAP)48 (Fig. 2c). In the absence of paired gene expression measurements, single-cell chromatin data can be independently clustered using Signac and manually annotated, or multimodal integration can be used to annotate the cell types in an unsupervised analysis41. To assess the accuracy of multimodal integration, we treated the gene expression and DNA accessibility assays as separate experiments and performed cell type label transfer from the annotated scRNA-seq assay to the unannotated scATAC-seq assay using the previously developed Seurat v3 data integration methods41. This revealed an overall label transfer accuracy of 87.0% for high-resolution cell annotations (Supplementary Fig. 3a) or 92.5% for lower-resolution cell annotations, with incorrect predictions occurring mostly between highly similar cell types (Fig. 2d). Furthermore, incorrect predictions received lower prediction scores, allowing low-confidence predictions to be identified (Supplementary Fig. 3b).
