Applicability for vanilla differential expression testing

[Thapeachydude]

Description of feature

Hello,

I would have the perhaps naive question of whether your tool is also suited for “vanilla” differential gene expression testing between conditions of single-cell populations. For instance I care less about predicting the response of patients and more about understanding the response to drugs.
As we see quite strong “patient-effects” in the gene expression, I’ve this far mainly used good old edgeR with a block for patient (y ~ drug + patientID). This works quite well with a pseudo-bulk per cell type, but is of course only a very course resolution for single-cell data (for instance within a cell type you might see two different populations arising, which you don’t see this way). 

I’ve tried more single-celly tools like lemur or contrastiveVI, but unfortunately, both tools produce latent embeddings that don’t respect core biological features (such diagnosis etc, therefore likely introduce strong artifacts). 

Any insights would be much appreciated : ) 
Cheers!

[MUCDK]

Hi @Thapeachydude ,

This is an interesting question. In general, I do believe this is possible, however, you would have to ensure that your encoder-decoder (e.g. PCA or VAE) is sufficiently powerful. However, I think this requires quite a bit of verification, so I don't think we should directly implement it. In general, we do have evidence in this work and other related work, e.g. https://arxiv.org/abs/2407.11734, that flow matching is able to capture nuances in heterogeneous distributions.

TLDR: I think it's a feasible approach, but requires some validation.