How can the cell line contribution be estimated from RNASeq data?

Question

Using a laser-capture microdissection of cells a group of cells stained with the marker of interest was sequenced. In another cohort of patients (this is all human liver tissue) the whole tissue was sequenced (RNA-seq in both cases)

Can I estimate the contribution of the cells marked in the whole liver ("weight of these cells" in the liver in words of my PI)?

My gut feeling is that it can't be done this way, it would require both single cell sequencing and whole tissue sequencing to estimate the contribution of each cell line. But perhaps there is a tool that given the cell lines or the main expression of other cells lines it can be compared to using GSVA or some similar tool.

Answer 1

There are couple of computational methods which try to do this (I never used them, so no experience):

1. CellMix, based on sets of marker gene lists
2. Subset Prediction from Enrichment Correlation, which is based on correlations with subset-specific genes across a set of samples.
3. Cell type enrichment, which uses our highly expressed, cell specific gene database
4. Cell type-specific significance analysis using differential gene expression for each cell type

You might have to get some reference expression levels from public databases or papers for some of the methods.

One thing to keep in mind: you cannot really compute the cells proportion, only RNA proportion. If you have a good reason to assume that RNA quantity per cell is very similar, this is a good proxy for the cells proportion in a tissue.

Answer 2

Cell deconvolution is mentioned in this Biostars post, which mentions CIBERSORT for immune cell mixes, and the Bioconductor package DeconRNASeq.

As far as I'm aware, it is only possible at best to get proportional representation for transcript expression from standard high-throughput sequencing results, because the sequencers and sample preparation workflow are designed in such a way that the same number of reads are output regardless of the input amount.

Answer 3

You can download public single-cell RNA-seq data or bulk RNA-seq for purified primary cells for reference and then use CIBERSORT or MuSiC to deconvolve. One thing to notice: to make bulk RNA-seq and single-cell reference compatible, TPM from RNA-seq can be compared with CPM in 3' (10X,Drop-seq etc.) single-cell data and TPM in full-length single-cell data (SMARTseq etc.)

Answer 4

Actually a month ago, I was searching for the same question and I found a few packages that can be used. First of all the functionality of them can be cell proportion estimation, cell specific gene expression estimation, or both of these functions. Secondly, their primary input, as expected, is bulk RNA-seq data and some also want a single cell RNA-seq data as secondary input. The most popular one is the child of famous cibersort, CIBERSORTX. I went through their website and read the tutorials. They have several example datasets that can be used for learning their platform. However, the two biggest flaws were that they haven't shared their source code which means we can't reproduce the result on our own systems and we have to upload all of our data to Stanford servers (security concerns!). The second problem is that there is no API access to their server and everything should be done online which means very low compatibility with other tools.

The idea, using deconvolution and deep learning to infer subpopulation of cells from bulk RNA-seq, is great. So, I searched to see if others have also done something similar and maybe even better. I found three other tools:

1. MuSiC: https://www.nature.com/articles/s41467-018-08023-x
   Published in nature communication and a few months older than cibersortx. I haven't studied their algorithm yet and so I can't say if their method is better than cibersortx or not. However, the part that makes the difference is that all of their works, including their source code, is open-source. So, I can access their code and we don't need to upload our unpublished data into some not safe online server (unlike cibersortx). The second good thing is that it is implemented in R language which I use for most of my RNA-seq data analysis, so it is easily compatible with other tools.

2. Scaden : https://advances.sciencemag.org/content/6/30/eaba2619.full This one is also open-source and accessible easily. It is written in Python and runs in the shell. So, it is somewhat compatible with other tools (not as compatible as MuSic). We again can process our data locally and don't need to upload our data to any servers (which is good). They claim that their method has a better correlation with the real scRNA-seq data than both cibersortx and MuSic (not always!!). Running their tools seems easy based on their tutorials (I haven't run it on my system yet!). However, their tutorials seem to be a little confusing and it seems like it is still in pre-release version and they have bugs!!

3. CDSeq: https://github.com/kkang7/CDSeq_R_Package Finally, I found the CDSeq package which doesn't need any scRNA-seq input to find single cell proportions and cell-specific gene expression in bulk data. It just gets your buld counts as input and info about very few parameters and does the whole thing automatically, and suprisingly good, I checked! If you could find gene expression pattern of your specific cells and also use it as an input, it is optional, it will give you even a better estimation.