Has anyone used singleR to predict the cell types of clusters in single cell? The algorithm predicts the cell type of each cell individually, but I was wondering if people have used it to predict what cell type each cluster represents? If so, can someone share the code?
The biggest advantage of the cluster-based approach is that it is faster. It is also a natural next step if you already have the clusters and you just want a direct answer to what they (probably) are. However, I prefer using the per-cell annotation where it is feasible, for various reasons:
It doesn't require clusters. This allows you to characterize the subpopulations in your data without doing the whole song and dance of normalization, feature selection, etc. I mean, you might end up doing these anyway, but the labels (and thus the downstream analysis) won't depend on the choices you make during those steps, which means that further downstream analyses on the labels are more stable.
It provides a better feel of the reliability of the inferred cluster-level cell type. Inconsistencies between the clusters and the assignments can manifest as clusters containing many different labels across its member cells; this would not be available if the cluster was aggregated into a single profile for labelling.
Of course, inconsistencies between the annotation and clustering can also be interesting, because it suggests that the clustering is capturing novel heterogeneity that was not present in the references. After all, if your dataset just recapitulates known cell types... sounds sort of boring.
Various methods are available to compare your de novo clustering to the reference annotations, see here.
