18 Automated annotation with SingleR
Automated cell type labels are assigned using SingleR with the Human Primary Cell Atlas as the reference dataset. SingleR uses a correlation-based approach to compare each cell’s expression profile against bulk reference profiles and assigns the label of the most similar reference cell type. The labels are added to the Seurat object metadata and visualised on the WNN UMAP.
library(SingleCellExperiment)
library(SingleR)
library(celldex)
sce <- as.SingleCellExperiment(combined)
ref.set <- celldex::HumanPrimaryCellAtlasData()
pred.cnts <- SingleR::SingleR(test = sce,
ref = ref.set,
labels = ref.set$label.main)
combined@meta.data$SingleR.labels <- pred.cnts$labels18.1 Cell type labels on UMAP
DimPlot(combined, reduction = 'integration.wnn', group.by = 'SingleR.labels')
18.2 Distribution of weight per cell type
The modality weights learned by FindMultiModalNeighbors reflect how much each
cell relies on RNA versus ATAC information when constructing its weighted
nearest-neighbor graph. Cells with noisy or low-information data in one
modality are down-weighted for that modality, so examining these weights
across cell types helps confirm that the WNN integration is behaving sensibly
rather than being dominated by one assay.
