I had a question on whether or not any method exists to work around an issue I have with some bulk RNA-Seq samples. 

Experimental background: I have a series of primary tissue samples in which we enriched for epithelial cells using mechanical dissociation techniques (scraping off with a scalpel). We knew this was a messy way of accessing the primary cells but the cell numbers/RNA quantity we get from the scrapings is already very low and cell viability with enzymatic dissociation is pretty bad. So we felt this was the best chance we had for getting any RNA samples at all. Further, when we checked, 90-95% of the cells we mechanically dissociate are positive by flow cytometry for epithelial markers suggesting we actually get a decent purity this way. However, we ultimately want to compare these epithelial primary tissue scrapings to in vitro cell lines we grew from these same primary scrapings. However, the in vitro cell lines are 100% epithelial cells as the culture conditions are selective for epithelial cells. 

Problem: The issue we are having is that when we compare the primary tissue epithelial scrapings to the in vitro cell lines, we see immune cell/blood cell specific signatures showing up in the top DEGs and in GO analyses. There are also some epithelial specific genes showing up but they are largely masked by the immune cell signatures which are an "all vs none" effect essentially between the tissue and cell lines.

Does anyone know of a method or creative idea on how we could deconvolute the bulk RNA-seq data to remove the immune cell signatures and focus on the epithelial signature to see how it changes once the cells are put in vitro? Is there any bioinformatics technique out there for something like this? Or is it a lost cause and we just should acknowledge that the tissues obviously are not perfectly pure epithelium and that some of what we see is from the contaminating cell types?

I tried to do some digging around on deconvolution techniques and that seems like a possibility, but I could not find a consensus on whether it could be applied to something like this or RNA-seq data at all.

Thanks!!