Deconvolution methods on RNAseq data

Question

[Ron]

Hi all,

This question is similar to this post. C: Deconvolution Methods on RNA-Seq Data (Mixed cell types)

I have been using CIBERSORT and DeconRNAseq, but both of them give different results.

1. Does anyone know of other cell type signatures besides LM22 from CIBERSORT which has only 22 cell types? Signatures that could include fibroblasts etc besides the 22 cells provided in LM22.

2. Also,are there any other tools for RNAseq besides CIBERSORT and DeconRNAseq?

3. Are there any softwares which consider other factors into account such as tumor purity before doing deconvolution?

Thanks,

Ron

[bruce.moran]

NB that the authors of CIBERSORT warn against using it on RNAseq (from their webpage):

LM22 was designed and validated on gene expression microarray data, specifically on the following platforms: Affy U133A/Plus2 and Illumina Expression BeadChip (HumanHT-12 v4). Users wishing to apply LM22 to other array platforms (e.g. Agilent single or two-color) or to RNA-Seq data should be forewarned that LM22 has not been validated for these platforms, and the results may not be reliable.

We are in the process of deriving an immune subset signature matrix for RNA-Seq and will inform registered users when it becomes available.

[Ron]

Are you still working on getting LM22 signature?

[bruce.moran]

This is a quote from the CIBERSORT webpage, so I am not generating an RNAseq-based LM22 signature myself, no.

Answer 1

[colonppg]

In my opinion the cybersort is not really reliable, this is particularly true for those cell types of low abundance:

essentially we want to use a gene, or a set of genes to represent the abundancy of a particular cell type, however, for the result to be reliable, those genes should be:

1. Have very high expression in the particular cell type it represents
2. No expression in other cell types

missing one of the two the result will be a joke... if you check their signature on reliable datasets, you will figure out it is a joke...

actually, I have been eyeing the whole cybersort and other "signature" methods as a joke as they just help the academics to publish papers based on shitty results without solving any real problem

[igor]

There is also a good Twitter thread that echoes much of this:

Anybody working on benchmarking methods for cell types deconvolution from bulk tumor RNASeq? #Cibersort #EPIC #deconRNASeq #infino #xCell (I’m quite concerned by the differences I see. Anybody else?)

— Nicolas Robine (@notSoJunkDNA) December 13, 2017

Answer 2

[shauty]

Hi, I just stumbled across this http://xcell.ucsf.edu/ . What's your experience with CIBERSROT? When I tried CIBERSORT I got only 16 results with p-value<0.05 from 75 cancer RNA seq samples.