If we use NMF for clustering single-cell, what are the k factors represents?

Question

[Raz]

I would like to use nonnegative matrix factorisation for clustering single cell. But, I am confused with what could be the representation of each k factors. I would like to cluster single-cell gene expression matrix (genes*cells) to detect cell subpopulations. Then let's suppose with consider k factors, what each k factor represents?

Answer 1

[Igor]

With NMF, you hope that the reconstructed basis represents some meaningful source of variability.

However, you need to understand that NMF is simply a mathematical procedure. It doesn't care whether you factorise single-cell matrix, or image, or text document. So the factor interpretation is up to you. It could be that some factor represents a particular transcriptional program, or it can be purely technical. It's the same as with clustering: you hope that clusters represent cell types. But that depends on how many clusters you have and how clean your data is. For example, cell might be clustered together purely based on high or low UMI counts.

I'd suggest looking at approaches such as MOFA(and MOFA+). That can give you an idea how to interpret your results.