Good afternoon,

I just started working with Single cell RNA Seq and I am trying to understand what is an average / possible number of cells per sample.

I am looking at a dataset of 7 patients with kidney tumour samples (please find attached). Would these number of cells be possible? I basically want to double check whether I didnt do any mistakes in my pre-processing.

Thank youenter image description here


1 Answer 1


It depends on many factors, both biological and technical, and it's impossible to answer in general.

The starting question is which technology was used, and how many cells were used in input. For example, in a SMART-Seq experiment with one cell per well (in a multi-well plate), you should know how many cells you have (one per well). For a 10x experiment, the person running the experiment should have measured the cell concentration in the suspension before loading the chip, and targeted a particular number of cells (typically 10,000). In a combinatorial barcoding approach (e.g. Parse), you'd also have a measurement beforehand so as to load a particular number of cells per well. So, you should already have an idea of the total number of cells.

Then, the first step of the analysis is typically some QC. In barcode-based approaches (including 10x), you'd make a knee-plot to estimate the actual number of sequenced cells in your sample.

In the clustering and annotation steps, you'd start distinguishing and annotating groups of cells, which hopefully do correspond to cell types. The algorithms are not perfect, so the clusters don't necessarily match cell types perfectly. For example, a rare cell type will easily be lumped together with another, similar, cell type. For a well-represented cell type, the algorithms have more chance to capture subtypes or cell states.

And of course there is the biology, and your knowledge of it. How many astrocytes are there in the kidney? I assume any number above 0 is worrisome. I'd expect the cell types in the data to roughly reflect the cell types in the tissue (but not exactly, there can be all kind of bias).

There is a question of depth: if you have few cells in a cluster, and few reads per cell, the gene expression data in that cluster is not very reliable. The actual number depends on the technology and read depth. I would say for 10x I would prefer at least 100 cells, and would not trust a cluster with <10 cells.

So for your question, you need to use your knowledge of the kidney and of the experimental design to decide. You can also compare to published scRNA-Seq experiments of the same tissue. You can also compare between patients to see if there is something troubling (for example, why does patient 1 have epithelial cells but no adipocytes, and opposite for patient 7? It might not be a problem, but only with context can you tell). And importantly, you do need to check that the annotation does match known gene expression patterns.

  • $\begingroup$ Dear Alexlok, Thank you so much for your detailed and very helpful answer, I will save it in my notes :). I have read a couple of times now the sentence "use your knowledge of biology" e.g. the found cell clusters. I definitely need to increase my cell biology knowlege here I think.... And yes also consider the experimental design more. $\endgroup$
    – Bine
    Jan 12 at 12:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.