Clustering resolution single-cell RNA-seq

Question

[Lucy]

Hi,

I have a question about how to decide on the clustering resolution to use for single-cell RNA-sequencing datasets that contain multiple activation states rather than discrete cell populations.

I have 10x Genomics scRNA-seq data from TCR-activated memory T cells from three donors (~2000-3000 cells/donor). When clustering is performed, the clusters represent distinct states of activation rather than discrete cell subsets. This means that it is not possible to use cell-specific marker genes to identify the optimal clustering resolution. I have also tried plotting clustering trees to find the optimal resolution, but the clustering is not particularly stable.

In this case, how would you decide on the clustering resolution to use and would you use the same clustering resolution for each of the three samples?

Many thanks,

Lucy

Answer 1

[jared.andrews07]

Personally, I would likely merge the samples from the all donors (optionally performing integration if it appears there are donor effects) first. Clustering is going to pull out differences no matter what, even if they're quite minor. In cases such as this, you need to utilize your biological expertise and what is being pulled out in each cluster to determine if they make any sense.

Additionally, heatmaps of the marker genes in each cluster can help you determine how well-separated the clusters are from each other. If multiple clusters highly express the marker genes for a given cluster, it may be necessary to reduce resolution. In short, there is no perfect way to determine this other than trying a handful and seeing which makes the most sense for your cells and experimental setup.

[e.r.zakiev]

good point about the visual inspection of the heatmap of the top expressed genes