If using singeR return more clusters than available ones should I do reclustering?

Question

[paria]

Hi all,

I did use a reference dataset to do automated annotation of my dataset. I do have roughly 20 clusters. But running singleR it annotated my dataset with more sub clusters. I was wondering should I redo my clustering and increase my clustering parameters or I could use the annotation of singleR for my dataset and do the DEG analysis?

Any comment would be appreciated

Thanks,
Paria

[jv]

But running singleR it annotated my dataset with more sub clusters

what do you mean by this? Are you saying that within one of your 20 clusters you had groups of cells from more than one cluster from the reference data set?

Answer 1

[jared.andrews07]

That is not how SingleR works - it is doing no clustering. I presume you mean that you're getting more than 20 labels back, which is not surprising since SingleR annotates on a cell by cell basis by default. If you want it to slap a label on each cluster, you need to pass them to the SingleR() call with clusters = a_vector_of_your_clusters. This is unsurprisingly faster to compute and easier to interpret, but it's making the (rather large) assumption that each cluster corresponds to exactly one cell type in the reference. If the cluster is actually a mixture of cell types, this will not be reflected in the output returned. The default per-cell method is safer and provides more information about the ambiguity of the annotations.

You can still label clusters by the majority of labels in that cluster after the fact (e.g. if a cluster contains 95% cells labeled as T cells, it's probably safe to label the entire cluster as T cells).

Regardless, you should not blindly use the results from SingleR (or any automated annotation method) without looking at the results to ensure they make sense. I'd recommend taking a look through the relevant diagnostics chapter in the SingleR book.