Mean-variance relationship in RNA-seq

[Slacanch]

hey all,
i am plagued by a problem regarding the mean-variance relationship in RNAseq experiments, hopefully you can help me with this.

according to most analysis tools i've used, rna-seq data can be modelled as a negative binomial distribution, which would be a poisson distribution with some more dispersion added, in order to account for biological variability.

In poisson distributions, the variance is equal to the mean. meaning that in a negative binomial, the variance is going to be bigger than the mean.

this implies that the larger the mean, the larger the variance.

my question arises when i'm faced with "mean-dispersion plots" in which the dispersion of genes is plotted in relationship to their mean read-count (usually in log2).
like in this example:

http://bioinf.wehi.edu.au/RNAseqCaseStudy/voom_mean_variance.png

if the variance increases proportionally with the mean (or more in the case of the negative binomial), how come the genes with the smallest mean count are always the ones that show highest dispersion in these sort of plots?

Thanks a lot

[ronaldrcutler]

Bump.

I am also curious about this too. It has been observed that the mean-varience relationship is approximately quadratic, while I also observe the dispersion parameter increase in genes with low counts.