Question

Deconvolution with CIBERSORT: input data

1

Entering edit mode

Matan G. ▴ 60

@matan-g-22483

Last seen 3.3 years ago

Hi,

I'd like to perform a deconvolution on my bulk RNA seq data. I'm testing CIBERSORT as an option.

Is the VST transformation on raw counts appropriate for use as CIBERSORT input? It is said that the data used as input should not be log-space normalized
If VST is not appropriate how can I actually normalize the data appropriately?
Since it is declared by CIBERSORT it seems that it is more appropriate for Chip data rather than RNAseq. Does anyone know an validated alternative (maybe a good R package)

Best

normalization r vst cibersort • 9.4k views

ADD COMMENT • link updated 4.3 years ago by Kevin Blighe ★ 4.0k • written 4.3 years ago by Matan G. ▴ 60

score 0 · Answer 1 · 2020-08-06

0

Entering edit mode

Kevin Blighe ★ 4.0k

@kevin

Last seen 16 days ago

Republic of Ireland

It seems that CIBERSORT prefers the input from RNA-seq to be normalised and follow a Poisson / Negative Binomial. So, please use the 'normalised' counts, i.e., not transformed via logCPM(), rlog(), or vst()

Regarding your point #3, I am not sure that it is true (?) - can you share the resource where you have read this? CIBERSORT has been around a long time and was initially developed on cDNA microarray data.

Kevin

ADD COMMENT • link 4.3 years ago Kevin Blighe ★ 4.0k

0

Entering edit mode

Hi, thank you for answering.

Are you familiar with an R package that can help with normalization methods as the one you suggested?

Regarding my comment on CIBERSORT RNA seq deconvolution, it is taken from the FAQs section in the official website. I'm attaching both a screenshot and a ref links.

Do you recommend CIBERSORT for RNA seq bulk data? Do recommend to try any other source for this application?

Best

Refs: https://ibb.co/tczTQ57 https://cibersort.stanford.edu/faq.php see "Can I use LM22 with data from other platforms, such as RNA-Seq? "

ADD REPLY • link 4.3 years ago Matan G. ▴ 60

0

Entering edit mode

Thanks, I can now [slightly] better interpret what they mean. They indicate that, if you use RNA-seq data, you should disable quantile normalisation in their options; however, it is not clear if they expect you to use RNA-seq normalised counts or normalised+transformed (vst, rlog, logCPM) expression levels. For full clarity, it may be better to contact the CIBERSORT group.

Fundamentally, though, keep in mind that CIBERSORT was designed on Affymetrix and Illumina microarray data, leading to the next point: Methods to predict cell-types in bulk RNA-seq data are still coming out. I believe there are a few post on Biostars about this.

Note also that this forum is for problems related to R / Bioconductor packages.

ADD REPLY • link 4.3 years ago Kevin Blighe ★ 4.0k

1

Entering edit mode

Hi Kevin,

Thank you for your answer. It has helped me.

I'm aware of the fact that this forum is less suitable thank you for pointing it out.

All the best

ADD REPLY • link 4.3 years ago Matan G. ▴ 60