Copy Number Variation from paired end RNA-Seq data

Question

[ChIP]

Hi!

I want to know, is it possible to identify copy number variations from paired end RNA-seq data? if yes , how? and which tools to use?

What I have done till now is:

a) Executed mpileup on the bam file of paired end RNA-seq file.

b) Next I wanted to use varscan copynumber, but it requires normal_pileup and tumour pile-up. I only have pileup on my experimental data. Hence I cannot move forward or am I missing a trick here?

Please help

Thank you

[igor]

I just realized I started another thread with some options listed: Detecting copy number alterations based on RNA-seq data

Answer 1

[Christian]

I have seen papers calling CNVs from RNA-seq data, albeit only the very large ones (chromosome arm-sized).

For example, have a look at Figure 1C and accompanying methods in this paper:

Patel et al. (2014) Single-cell RNA-seq highlights intratumoral heterogeneity in primary glioblastoma

Answer 2

[Devon Ryan]

The only way to see CNVs with RNAseq is if some of the copies result in fusion events with other genes, thus resulting in one read in a pair aligning to gene A and the other to gene B. Aside from that, one can't look at CNVs with RNAseq.

Even with a matched control sample, you still can't use the method you outlined to call copy number. It's impossible to discriminate between differential expression and an underlying copy number variation with RNAseq data. This is a fundamental limitation. If you want to look at CNVs, you need to sequence DNA.

[Christian]

I wouldn't say it is impossible, just limited to very large CNVs due to the high coverage variability introduced by differential expression (see my answer and the referenced paper below).

[Devon Ryan]

True, if the CNV is absolutely huge like that then it can work.

[ChIP]

Will WGBS can help in detecting CNVs?

[Devon Ryan]

Yes, you can call CNVs from WGBS data (I've never seen anyone do it, but there's no reason you couldn't). You could use the mpileup->varscan copy number method with that.

[Devon Ryan]

I take that back, while you can use WGBS data to call CNVs, varscan still needs a control dataset. You might be able to find something close enough in a public dataset, if you're lucky. Otherwise, there ends up being no great way to call CNVs. If one were to use the naive method of just going by depth, then any bias introduced by mappability, library construction, etc. could seriously skew the results. Having said that, if you're going to validate things anyway and you can tolerate a moderately high false-positive rate, then perhaps this can suffice.

[ChIP]

Thank's, will keep your points in mind.

[Devon Ryan]

Best of luck with this. Feel encouraged to leave a comment when you have the results. I'm curious to hear how well this works in practice.

[ChIP]

Thank's again. I have no clue how well this will work as I do not have control datasets, but we will validate some of our calls so lets see. Will keep you posted.