I am trying to understand library normalization in DESeq2. I would like to ask the following: I know that some samples have been run 15 cycles and some others 20, can I give this information to DESeq2, would it be useful? I mean, I know that DESEQ2 uses an algorithm to normalize library size, but since I have this piece of information, I wonder if it would be useful to provide it somehow to DESEQ2. If it is not, would it make sense to incorporate this information in the design matrix, something like: design(~cycle+condition)? Thank you a lot!

  • 1
    $\begingroup$ What does "circle" mean in this context? Do you mean some number of PCR cycles? $\endgroup$ – swbarnes2 Sep 30 '20 at 19:27
  • $\begingroup$ I am sorry, I meant cycles (I edited my question)! $\endgroup$ – marilu Sep 30 '20 at 19:49
  • $\begingroup$ what is dds??? is it possible to ask 1 question $\endgroup$ – StupidWolf Sep 30 '20 at 19:51
  • 1
    $\begingroup$ You can include it in the design matrix, but you need to be careful. If you use it as a continuous variable, then you are trying to find a linear relationship between counts and the cycle number. This may not be the case and in same cases you actually introduce more noise into the model $\endgroup$ – StupidWolf Sep 30 '20 at 21:37
  • 1
    $\begingroup$ I think you are trying to make a model more complicated than it should be $\endgroup$ – StupidWolf Sep 30 '20 at 21:38

You can include that info in the design, though you'd have to decide whether to include it as a number of just a factor. I've never heard of anyone including PCR cycles. I've heard of people including RIN scores, though I've never done it.

If the PCA looks alright, I wouldn't worry about. I've got tons of examples of the lab people saying "I did a few more cycles on these few samples" and then you can't tell any difference in the PCA. I think needing to do more cycles likely has more to do with variance in the prep than a real biological difference.

  • $\begingroup$ Thank you a lot for your reply!The problem is that PCA does not look good and I have observed that the samples with 15 cycles tend to cluster. I have tried to use removeBatchEffect from limma (I made a design matrix with cycle as a factor).then the pca got better. In the downstream analysis I used the new design matrix with cycle like in the question, and then the differentially expressed genes got dramatically reduced...(from 750 to 100).Any thoughts? $\endgroup$ – marilu Sep 30 '20 at 20:05
  • 1
    $\begingroup$ It is possible to have some real DE genes even if your PCA looks discouraging. 100 is not a lot, but maybe that's realistic if your condition doesn't split nicely along your first few PCs. You might play around with making cycles a factor vs a number. But if it's right that the number of cycles is really driving some difference, or representative of some real difference, accounting for it in the design will attenuate those differences. $\endgroup$ – swbarnes2 Sep 30 '20 at 20:17
  • $\begingroup$ Thank you so much for your reply, it really helps. One more question, I have observed that when I use vsd instead of rlog for PCA pot, PCA plot looks better. However, my samples are less than 30 (in the tutorial, rld is recommended especially for samples less than 30). Any ideas? Thank you in advance. $\endgroup$ – marilu Sep 30 '20 at 20:22
  • $\begingroup$ I think the author recommends vst for everything now. $\endgroup$ – swbarnes2 Oct 1 '20 at 21:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.