What would you do with a richer, less manual way to annotate cell types? 

Single Cell Portal sent a dataset to Cellarium Cell Annotation Service (CAS) to see what would come out of such a tool. Let us know what you would do in our CAS survey!

Here are the vanilla clustering results 10x Genomics supplied with this small PBMC dataset:

Cellarium CAS generates the most granular cell type annotations that meet a given probability threshold. Pictured here are cell type annotations from CAS where the annotation probability is at least 65% (Annotation: high_conf_CAS_set_1_cell_type) 

Corresponding confidence scores for each annotation visualized with the high_conf_CAS_set_1_probability annotation.

You can also obtain the next most likely annotation at the same threshold (Set Annotation dropdown to high_conf_CAS_set_2_cell_type, high_conf_CAS_set_2_probability)

Setting the probability threshold to 10% reveals matches to more granular cell types. These more granular labels may be less plentiful in the reference dataset, resulting in lower probability matches. (granular_CAS_set_1_cell_type and granular_CAS_set_1_probability)

Looking across annotation sets at the original clusters 3, 4, 10, 11 and 13 (see clusters in oval) CAS results imply these are not distinct cell types because they are uniformly annotated with a single cell type annotation across the CAS-assigned annotation sets -- the most specific cell type being "classical monocyte".

By contrast, CAS annotations of clusters 2, 9 and 18 (see clusters in circle) show more heterogeneity than the original clustering would suggest.

In Single Cell Portal, you can use the legend labels to visualize individual cell type annotations separately. Notice that cluster 7 (see square in figure above) has both "Natural killer cell" and "Mature NK T cell" annotations. Toggling these cell type annotations in the legend allows us to see that these are overlapping cell type annotations.