Check if RNASeq count data follow NB or Poisson distributions

Question

[debitboro]

Hi Biostars,

I have used HTSeq to generate the following table for counting reads per gene per sample (I have 12 biological replicates):

ENSG00000000003  0  0  5  7   0  0  0  0   0   0  12   0
ENSG00000000005  0  0  3  2   0  0  0  0   0   2   4   0
ENSG00000000419  2  2  3  5  18 20  0  2   2   3  13  32
ENSG00000000457 15  6 11  7 129 21  8 90  41  97 129 104
ENSG00000000460  6  2  9  5  62 12  3 30  21  61  78  62
ENSG00000000938  0  0  5  0  16  3  0 16   7  25  32   5
...
...

My data are paired-end RNASeq data. Now I want to check if my count data follow NB or Poisson distributions. What is the recommended way to perform this ?

I appreciate you help.

Answer 1

[Devon Ryan]

Plot the variance as a function of mean (use normalized counts). If there's a linear relationship (there won't be unless you're working on a cell line or something simple like that) then it's Poisson.

[Michael Love]

To see if gene counts from technical replicates are well approximated by Poisson, I've tried looking at the SEQC technical replicates, using the Bioconductor seqc package. Poisson was a good fit for most genes. It's a bit tricky because of differences in library size across samples, so I used the expected value for a gene x sample as the rate of the Poisson and looked at the distribution of cdf(count). This was nearly uniform.