RNA-seq counts distribution

Question

[ecj]

Hi there,

I am quite new in the bioinformatics world and I would appreciate a bit of help, please. I would like to have an explanation about the counts distribution of a RNA-seq analysis. What I have studied and understood, is that they typically follow a negative-binomial distribution and according to that, the calculations of the DESeq function are appropiate for the analysis. My question is: if I do a "manual" filtering and remove the genes with a base mean under the threshold I choose (this filter out around a 70% of genes), when I plot the counts distribution of the remainin genes, they do not have this negative-binomial distribution as before. Instead, their distribution is more similar to a normal distribution. Then I would like to know, is this affecting the analysis? Should I consider other different ways to analyse now the data?

Thank you in advance.

Answer 1

[i.sudbery]

In RNA-seq, the inter-replicate variation is described by a negative-binomial distribution, not the inter-gene variance. When you are doing DE, you care about the variation between replicates/conditions, not between genes.

The distribution of mean TPMs between genes is generally thought to be a bimodal normal. See Hebenstreit, D., et al (2011) Molecular Systems Biology, 7, art. no. 497, .

Thus, if you off the bottom of the distribution, you are likely to be left with a normal.

Answer 2

[Istvan Albert]

When you remove the majority of your data - as you noted - it may fundamentally alter the distribution.

It would be not surprising if the removal radically influences the results you get. Whether or not the results still make sense is impossible to tell from our vantage points (as commenters on the forum) you would need to closely investigate the processes and results you get, contrasting it to not applying the removal for example.