When we have three groups samples (A,B,C) with 25000 genes and the main interest is A vs B, should we limit samples only A and B for normalization to perform DEG analysis? Or better to include all samples (A,B, and C) for normalization and specify the comparison group within three groups?

res <- results(dds, contrast=c("tissu_type","A","B"))
  • 1
    $\begingroup$ No gold standard for this. If you probably are going to need C at some point then include it. If not then leave it out. If C is very different from A and B it might negatively influence normalization and dispersion estimation. Depends on your analysis goal. I would probably include it because if you ever need C you have to repeat the analysis and results from A/B would change. $\endgroup$
    – ATpoint
    Jun 14 '20 at 9:39
  • $\begingroup$ I guess it makes more sense to include A and B. Imagine you are writing a paper, and it simply does not make sense if you include C, which you never talk about in the paper. But there is no absolutely correct answer for this. I also believe that the results are going to be very similar. If the results using two methods are so different, then there might be something wrong. $\endgroup$
    – Phoenix Mu
    Jun 14 '20 at 19:31
  • $\begingroup$ Thank you so much for your all comments! $\endgroup$
    – user224050
    Jun 14 '20 at 22:17

As explained on the comments, having some other samples might distort your estimations, add an outlier or something alike.

The advantage of having all the three (or all) groups of samples before testing differences, because usually methods like limma, DESeq2 and others uses the information of variation of the genes across all samples to update their prior probabilities and the variation of the genes on the statistical model used to compare between groups.

If the method and your model are robust use all the three groups (generally the methods used are robust, but sometimes the models are not). By robust I mean that it doesn't miss any know (and available) source of variation. Otherwise the comparison might be comparing only the samples not the effect of the condition on the samples. If you only compare the samples it would be equivalent to use a t.test to the limma methods (not completely sure on the DESeq2 case). Or explained differently, you wouldn't take advantage of the statistical improvements implemented on these methods

Usually the reasons why these groups are sequenced on the same batch is because they are related (or to add some known ground truth), so you can release it with the data when communicating your results but you may omit that you used this third group on the methods section.


Plot out the PCA of your samples. If, say, PC1 splits the C's away from the others, and PC1 is more than, say, 50% of the variance, or if the C samples are a lot more spread out than the A and B samples, keep it apart. Otherwise, do them all together.

If the C's aren't too different from A and B, the math will benefit from including them.


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.