How would you annotate single-cell clusters with cell types and judge how confident the labels are?

Why

Annotation turns clusters into biology, and the Bioconductor book calls this interpretation step arguably the most challenging task in single-cell analysis. Two approaches complement each other. Reference methods assign each cell the label of the reference samples it most resembles.

SingleR uses Spearman correlations on marker genes between pairs of labels and then fine-tunes among the top candidates. Its reliability depends on the reference, including the expertise of whoever labeled it. Marker-based annotation looks at the genes that distinguish each cluster from the others and matches them to known biology, such as CD3E for T cells or MS4A1 for B cells.

Comparing the two is informative. When a cluster gets a consistent reference label and expresses the expected markers, the label is solid. When they disagree, the book notes the discrepancy can reveal a new factor of variation, such as a state the reference never contained.

A reference cannot assign a label it lacks, so an unfamiliar population is forced onto its nearest match. Confidence also falls for doublets, stressed cells and clusters defined by a few genes.