Hello,

We ran DESeq2 on an experiment with several samples (each sample in triplicates). The samples differ dramatically. In one of the pair wise comparisons, we have ~4200 differentially expressed genes (with fold change above 2 or below -2, p-adj<0.05), out of ~19,000 detected genes (human genome).
DESeq normalization is based on the assumption that most genes are not differentially expressed.
Do you think that it is valid to use this normalization in this case?

Thank you

Thanks a lot for the answer! I am pasting below an image of the MA plot (I hope it is OK to add an image and it will appear OK).
I think it is OK. Do you think that the normalization is OK? (DESeq was run with beta prior False).

 

That's not exactly the assumption. It's that the median ratio captures the sequencing depth differences. So there are many scenarios where you could have many DEG and still have good normalization using the standard size factor estimates.

Take a look at plotMA(res) to see if there are genes with large mean count falling along the x axis.

Thanks a lot for the answer!
Here is a link to the MA plot:
https://weizmann.box.com/s/dtd75ielrfxxf6vvek3ziuk6tel27a8p

I think it looks OK. Do you think that the normalization is OK? (DESeq2 was run with beta prior False).

Thanks!
Here is a link to the zoomed out MAplot:

https://weizmann.box.com/s/vx9djse9s4fbgnwr373h03kbby0opfc9

I have seen large number of differentially expressed genes if 

1) Experimental design, for example there is a batch effect that is confounded with the biological effect, for example all controls were performed in one batch and all the treatment were performed in a separate batch. In this case the differentially expressed genes could be due to the biological or batch effect and there is no way to determine which is which.  

2) Replicate issues. Are your replicates technical or biological

3) Its true (rarer) ! Maybe you really have a very large biological effect.  For example estrogen receptor positive v negative breast cancer is very different, by about 2,000 genes. Or you are disrupting a fundamental cell pathway, for example an key gene in transcription (polII) and thus lots of genes are affected.  Have a look at the genes that are differentially regulated, do they make sense in your cell system.  

Aedin

Thanks a lot for the response!
The replicates are biological and not technical (and they cluster very nicely). The samples are very different biologically, so we do expect to have lots of changes. I am just wondering whether in such a case DESeq's normalization is valid.