This is a test version of Biostars. For the public version, visit https://www.biostars.org.
Differential expression analysis in DESeq2 for an experiment with multiple variable

Dear all, I am approaching a kinda new experiment and wondering what is the best way to construct my design, to plug into DESeq2, in order to best address my question of interest.

  • First we have, 10 replicates of mouse cells from tissue X taken at two different steps of their life cycle (let's say cycle A and cycle B). These cells were taken from lab 1. Samples were paired-end sequenced.
  • Second, 10 replicates of mouse cells from tissue Y taken at two different steps of their life cycle (again, cycle A and cycle B). These cells were taken from lab 2. Samples were single-end sequenced.

All of these samples are balanced and/or paired (?), meaning that each biological replicate is taken from the same mouse each time. For example, from mouse 1 were extracted cells at both cycle A and cycle B.

So, same cell types at two separate conditions but extracted from two different labs (in two different parts of the globe). I am relatively new to this kind of design formula. But, what I want to do is:

  • First, find differentially expressed genes within each lab (i.e. A vs B from lab 1; A vs B from lab 2)
  • Second, try to compare everything across each lab in all the possible combinations but also caring of the different above mentioned variables.

To sum up we have a four different variables

  1. cycle

  2. tissue

  3. replicate

  4. lab

Therefore I would suggest a design matrix like the following one: ~ tissue + tissue:cycle + replicate + batch + cycle

I am just throwing an idea, I am not 100% sure.

Thanks in advance

deseq2 rna-seq
• 1,357 views • link
updated by ATpoint ATpoint • written by Mozart Mozart

1 answer

Lab is fully confounded by tissue. That means any differences between tissue could in theory be 100% due to batch effects and probably most DEGs will be. You cannot correct for this because of the full confounding. Doing this contrast would not be wise in my opinion. There are limitations in data analysis, and comparing independent RNA-seq data, acting as if they came from one experiment, is one of it.

Replicate does not belong into the design formula. Please check the DESeq2 manual on how to indicate replicates in the colData.

First, find differentially expressed genes within each lab (i.e. A vs B from lab 1; A vs B from lab 2)

That is the only meaningful analysis you can do. Create two separate dds objects, one for each lab, and then perform standard analysis. No interaction terms etc., simply ~cycle.

Second, try to compare everything across each lab in all the possible combinations but also caring of the different above mentioned variables.

As I said above, I do not see which results beyond a reflection of the batch effects this would yield. What does lab mean here? Is it published data so samples are completely different or are those samples from e.g. two collaborators and they at least did all the library prep. etc exactly identical? As said in your question last week, a paired-end sample can always be treated as a single-end sample so this alone is not a problem, it is the library prep process which induces the major confounding.

• 0 views • link
written by ATpoint ATpoint

Thanks a lot for comments provided. These results are actually from two collaborators (don't think these results are published yet). But again, this is just an exercise of style for me as someone else is in charge for doing this analysis. Someone who I don't personally know. The latter generated some "interesting" results people want me to reproduce again.

I totally agree with you when suggesting that the best thing to do is actually creating two dds object in order to perform a standard analysis. But the idea of this collaboration rises for the precise attempt to have these two datasets merged together, for how wrong can be. I am sorry if it seems I want to insist on this but the thing is that sometimes is absolutely difficult to have a clear say when, for the sake of saving money from an experiment, you receive orders from the top regardless of the fact that these people have absolutely no idea what they are doing.

Therefore, in the light of what you said I think you can't suggest any viable alternatives to, at least, try and draw some information from this cross-comparisons, I guess? there is absolutely no way to attempt having a results? PS: I think I have done millions of time correctly but could you please confirm that replicates are explained in colData as 'sample' that would be very reassuring.

• 0 views • link
written by Mozart Mozart

Log in to answer this question.

More posts like this