Validating snRNA-seq cell type by correlating with other datasets

Question

[Ben]

Hi all,

I am re-analyzing data from a paper (paper 1) that finds cell type X in their snRNA-seq dataset. I want to distinguish between subtypes of cell type X (X1 and X2). I found another snRNA-seq paper (paper 2) in the same organism that makes this distinction between cell type X1 and X2. My goal is to sub cluster cell type X in paper 1 and then validate that these sub clusters are cell type X1 and X2 by correlating with paper 2's dataset.

My thinking right now is to average gene expression across X1 and X2 and then correlate the shared genes across datasets. Alternatively I could try to integrate paper 1's clusters into the UMAP space of paper 2 and see where they cluster?

I've tried the first approach (correlation of average gene expression) and the results were not promising: paper 1 X1 correlated better with paper 1 X2 than paper 2 X1. But part of me is not surprised at all. I am trying to differentiate between a quiescent and active state of a rare cell type. It makes sense to me that there is more variation across datasets than quiescent vs active cells. Is there any way around this?

What are best practices for validating specific cell types across datasets?

Thanks!

Answer 1

[ATpoint]

I would not do correlation analysis like this, I don't see how this makes any sense. Integration is possible but cumbersome. Instead, use SingleR (R/Bioconductor package). It uses a Spearman correlation-based approach but per-cell using a reference set of count matrix and markers (your dataset) and then predicts the best match for each celltype in a query dataset, based on your (or generally the reference) dataset. That, together with the scores of the prediction will help to tell whether a) the celltypes are even present and b) whether they somewhat aggregate in unique or dedicated clusters.