Did I understand CytoTrace kernel correct?

[dawe]

Hi all, I am looking at how CytoTrace kernel is implemented. Is it correct to assert that the ct_score is the cell-wise mean of the expression of the top 200 genes most correlate with the (cell-wise) total number of expressed genes?
If so, I can't understand what is the rationale behind the idea.
Also, would it be useful to allow parametrization of the top_n correlated genes?

[dawe]

Nevermind, I have thoroughly read the paper to get the rationale. Maybe it would be a good idea to state, somewhere in cellrank's documentation, that this method is particularly suited for developmental processes, and it won't work properly if there is not appreciable decrease of the number of expressed genes in the trajectory.

[dawe]

Hello @Marius1311, I have read the paper (not before reading the code, shame on me!). In the hematopoiesis example they still start from a stem cell and also the luminal/basal transition is quite extreme. I will open a PR as you suggest. If you don't mind I'd like to add a top_n option (default = 200). I'm considering the option to reverse the correlations order to take into account other kind of experiments in which the priming cell have an opposite phenotype. In this sense I'm thinking to scCUT&TAG experiments for close chromatin (e.g. H3K9me3 or H3K27me3).

[Marius1311]

Absolutely, please add PR for this and we'll be happy to review it!!

[Marius1311]

Sorry, I'm a bit out of the loop, @michalk8 already resolved this in #761.

[Marius1311]

Can you explain what you mean by "reversing the correlations"? Basically, the CytoTRACE kernel just calls the PseudotimeKernel under the hood, pretending that the CytoTRACE score is a pseudotime. The PseudotimeKernel is very general, it accepts any measure of cell state together with a KNN graph and computes a directed transition matrix from it. If there's an option to get such a measure from cell state, i.e. something that quantifies cell plasticity on a continious scale, from scCUT&TAG experiments, this would be super cool and I would be open to intergration in CellRank.

[dawe]

@Marius1311 ctk sorts correlations of gene expression with number of expressed genes from top to bottom and takes the top 200 mostly correlated (right?). I’m thinking to the possibility of taking the top N with most negative correlation.
BTW in our lab we are working on using cellrank on epigenome data, it would be great if any of the developers could chime in