Entering edit mode
Hello Community,
The paper of DESe1 says that the forumula (6) :
A parametric curve of the form (6) is fit by regressing the gene-wise dispersion estimates α_g^wi onto the means of the normalized counts, μ̄ i via a gamma-family GLM regression.
Wihtin the Source code of DESeq2 I found the following code:
# Estimate a parametric fit of dispersion to the mean intensity
parametricDispersionFit <- function( means, disps ) {
coefs <- c( .1, 1 )
iter <- 0
while(TRUE) {
residuals <- disps / ( coefs[1] + coefs[2] / means )
So am I right that one assumes something like that? Atr = a1 / m + a0 + error
Atr /( a1 / m + a0 ) = error /( a1 / m + a0 ) mit epsilon = error /( a1 / m + a0 )
Atr /( a1 / m + a0 ) = epsilon ~ Gamma(a0, a1)
Thank you in advance
Thank you very much for your very helpfull answers. Unfortunately I am stuck in the section of Dispersion Prior:
Additional file 1: Table S2 compares: - ψ1((m−p)/2) as an approximation of σ2lde with - the variance of logarithmic Cox–Reid adjusted dispersion. And says it is very similar.
As a result:
I do not really get why one substracts the squared logarithmic residuals from the trigamma approximated logarithmic dispersion estimators due to the fact that the logarithmic Cox–Reid adjusted dispersion are similar to the trigamma approximated ones. in order to obtain the prior variance of log(a_i).
I think I am missing something conceptionally - it would be great if you could give me a hint.
Thank you in advance.
The subtraction is from what you would expect with a hierarchical Normal. The variances of the Normals add, so to get a reasonable prior from the observed distribution, we subtract the expected sampling variance.
So I am wonderin what the hierarchical normal model looks like:
) log(ai^gw) - log(atr(\bar mui) ~ N(aigw, Slr)
) ai^gw ~ N(ai; sigma^i_lde)
) log(ai) ~ N(log(atr(\bar mui); Sigma^2d) as prior
It is really interesting for me what you did - but I am really stuck.
Thank you again.
Yes, the second two lines are the assumptions that motivate subtracting the two estimates.
Hey - Thank you again.
So I was trying to draw sthg like this:
https://pasteboard.co/IdKFNpx.png
Which is from: http://idiom.ucsd.edu/~rlevy/pmsltextbook/chapters/pmsl8.pdf
Does it make sense?
Thank you in advance
Yes, that kind of diagram makes sense.
=) nice thanks.
Further it would be great if you are so kind and comment my suggestion (the link) of how to derive the sigma^2d of the normal hierarchical.
This is beyond the scope of the support I can provide here. On the support site, I try to answer basic questions about how to use the software, or clarification questions like the above thread.
I understand that - thank your for your help.
Another question came to my mind.
Am I getting the cox-reid adjustment right. 1/2 log det xt w x One penalized values for alpha which have a lot of information on mu?
If so I do not get why this makes sense.
To emphasize on values of alpha coming more from the poisson distribution?
Thanks again
hi Philipp,
Again, I'm kind of pressed for time these days and need to reserve my support site allotment for pressing questions about usage of the software. I try to make myself very available for these questions, but I received for example 32 questions since Monday, so I have to be as efficient as possible. You can look over the Methods, and the open source code, and follow the references in the original paper to understand our use of the Cox-Reid adjustment, where we follow edgeR. I don't have time to check your notes here.
Hey Michael, Yes your support is great! I am sorry that I bothered you with my questions. All the best =).