I have a single-cell RNASeq sample, in which I'd like to identify latent variables (e.g. response to stress) that I think might be affecting the clustering.

The approach I was planning to use is to reduce the dimensionality of the dataset using one algorithm among:

  • PCA
  • ICA
  • NMF

Then identify the most important biological processes for every component (e.g. in each Principal Component), and then finally regress out the processes that might not be relevant to the analysis or that might affect the downstream clustering.

I have a few questions:

  • Are there issues in this approach?
  • Is there a collection of gene sets (similar to MSigDB) that can be used to identify noise latent variables?
  • Which algorithm (PCA, ICA, NMF, or others) is more suited to this kind of analysis?
  • $\begingroup$ ..... PCA + tSNE $\endgroup$ – Michael Mar 19 at 21:11
  • $\begingroup$ @Michael why tSNE? How do you use tSNE in such case? $\endgroup$ – gc5 Mar 19 at 22:22
  • $\begingroup$ Its brilliant.... :-) $\endgroup$ – Michael Mar 19 at 23:11
  • $\begingroup$ The larger question is how do you know you are regressing out the right stuff? $\endgroup$ – StupidWolf Mar 20 at 10:32
  • $\begingroup$ And if the so called latent variable is taking up so much of your variance... how can you be sure of the response of your interest $\endgroup$ – StupidWolf Mar 20 at 10:33

I think to identify latent variables, PCA is probably not going to work. NMF might be worth trying. You might want to check out a method called consensus NMF (cNMF) (https://elifesciences.org/articles/43803) published in eLife. The author described some latent variables corresponds to cell types, while others corresponds to cell-cycle or metabolic states of cells.

| improve this answer | |
  • $\begingroup$ Thanks! I will read it now! $\endgroup$ – gc5 Mar 23 at 19:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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