No tumour cells in clustering of tumor sample?!

Question

[Bine]

Dear all,

First of all, please apologise if this question is maybe a bit stupid but I just started working with single cell RNA seq & I am not a biologist.

I have performed clustering of single cells (via Seurat) and afterwards annotated the clusters (via SingleR). This was done for tumour samples. Now I was expecting something like in the picture attached, showing some tumour cells. However, I only get endothelial, t-cells, b-cells, macrophages and so on..

Can anyone explain me this?

Thank you!

[swbarnes2]

Can singleR add that particular annotation all by itself?

Answer 1

[jared.andrews07]

I expect whatever reference you used with SingleR did not have tumor cells included. Only labels for cell types included in the reference dataset will be assigned, so it's important that your reference contains all the cell types you expect. Otherwise, it'll just assign the closest match.

[Bine]

Thank you so much! That makes totally sense. Is there anywhere an overview (maybe even by cancer type, in my case kidney cancer) which reference datasets are appropriate to use? I find it a bit difficult to find information on this to be honest. jared.andrews07

[jared.andrews07]

Well, depending on what you're trying to do, automated annotation may be rather pointless other than for identifying cell types of infiltrating/contaminating normal cells in the tumor. The labels assigned to the tumor cells may or may not be of any real use.

For tumor cells themselves, if there exist single cell datasets of tumors with meaningful classifications or annotations of some variety (for instance, some of the brain tumors our group studies may have OPC-like or neuronal-like malignant cells), you could use those. You'll have to find them yourself for your tumor type of interest though.

[Bine]

jared.andrews07 thank you very much for your reply. Are there any specific requirements my reference dataset should fulfil in order to work reliable (e.g. number of samples). I dont find any specific details on this?

> ref- A numeric matrix of (usually log-transformed) expression values from a reference dataset, or a SummarizedExperiment object
> containing such a matrix; see trainSingleR for details.
> Alternatively, a list or List of SummarizedExperiment objects or numeric matrices containing multiple references. Row names may be
> different across entries but only the intersection will be used, see
> Details.

On the other hand are there any big scRNA annotation data collections to check before searching for datasets randomly. I found "scRNAseq" and "TabulaMurisData" packages however both dont contain my cancer of interest. But maybe there are more collections like this?

[jared.andrews07]

The reference dataset requirements are flexible. Generally, more samples/cells in them will probably be more reliable, but it really comes down to how accurate the annotations for them are. SingleR can take bulk or single-cell reference datasets. Given that most single cell datasets have dozens to thousands of cells for most populations represented, sample number is typically not a concern, per se. Then again, if the annotations in the reference are not accurate, that will carry over to the test dataset as well. It's good to look at some individual markers as sanity checks afterwards.

There are thousands of scRNA datasets at this point. I expect your cancer is not so rare as to have never been sequenced. I do not work on kidney, so I have no knowledge of kidney-specific datasets. There are lots of datasets in the scRNAseq package, though IIRC none of them contain tumor samples. celldex also has several datasets along with curated labels, but again, no tumor samples.

You should think carefully about what question you hope to answer from this and whether this is the appropriate process to do so.

[Bine]

Thank you very much for your feedback.