To assess multimodal label transfer accuracy, we treated the DNA accessibility and gene expression assays of the multiome dataset as though they were separate scATAC-seq and scRNA-seq experiments. We computed a gene activity assay for the scATAC-seq dataset by counting fragments overlapping the gene body and a 2-kb upstream region for each gene in each cell, using the GeneActivity function in Signac. We log-normalized the gene activity counts for the DNA accessibility assay using the NormalizeData function in Seurat41,42. We then identified anchor cells between the scATAC-seq and scRNA-seq datasets using canonical correlation analysis, with the function FindTransferAnchors in Seurat with the parameters, reduction = ‘cca’. Cell type labels were transferred from the scRNA-seq to scATAC-seq dataset using the TransferData function, with weight.reduction = query[[‘lsi’]] and dims = 2:30 to weight anchors based on nearest-neighbor distances in the LSI space. To evaluate the accuracy of multimodal label transfer, we counted the number of cells obtaining the correct predicted cell type label.
