Hello,

I have a few questions regarding the normalization performed by the DESeq2 package:

• I understand that the normalization is working under the assumption that most of the genes are not DE and only a small subset of genes are. I understand that this assumption is realistic in most experiments, but I would appreciate your input regardless the correctness of this assumption when performing analysis across different human tissues. Meaning, does it make sense to assume that most of the genes will have the same expression levels in different human tissues? or should I perform different normalization when dealing with this sort of data?
• In some RNA-Seq experiments we encounter situations in which some libraries are sequenced in a greater depth than others, with larger range of library sizes that we usually expect. In those cases, the size factor given to the deeply sequenced samples will be much higher than 1 and the factor assigned to the smaller samples will be much lower than 1. In those cases, is it better to subsample the larger libraries in order to prevent this wide range of size factors? 
  I am addressing this issue, since when a sample receives a very small size factor, it seems to me that we are artificially increasing the counts of all genes in this sample without any biological evidence to back that up. It seems to me less problematic to lower the counts of the bigger libraries, but the artificial addition to the small ones seems more of a problem to me. What do you think?

Thank you very much,

Olga. 

hi Olga

"normalization is working under the assumption that most of the genes are not DE and only a small subset of genes are."

It's not that the subset is small, but that the median ratio captures the scaling factor from the non-DE genes. You could have many DE genes, and with balance of up- and down-regulation, the median ratio would be fine. The best way to check if the assumption makes sense is an MA plot of the results table. For many tissue comparisons, one will see a bulk of genes on the x-axis, with DE genes appearing above and/or below. For experiments where all genes increase in expression across conditions, the median ratio method will not be able to capture this difference, but this is typically not the case for a tissue comparison, as there are many "housekeeping" genes with relatively similar expression pattern across tissues.

"is it better to subsample the larger libraries in order to prevent this wide range of size factors?"

No, subsampling generally reduces power, as we are throwing away observations. The model including a size factor parameter handles this.

"it seems to me that we are artificially increasing the counts of all genes in this sample without any biological evidence to back that up."

It's not increasing the count, but decreasing the expected count from the generalized linear model to match the counts observed. It's a subtle statistical point about modeling with parameters versus transformation of data. See the formula in the vignette in Section 4.1 The DESeq2 model. We have that E(K_ij) = s_j q_ij. So if we have two samples with the same q_ij, but sample A is given size factor s_A = 0.5 and sample B is given size factor s_B = 2.0, then the expected counts E(K_ij) are lower and higher for A and B respectively to match what we know about the sequencing depth of the experiments.