[BioC] DEXSeq dispersion trend estimation?

[Ryan]

Hi Alejandro,

Thanks for the clarification. I have another more speculative questions. 
It seems that if you just put a trend on all exons in one big set, you 
are ignoring the possibility that exon dispersions within a gene are 
more correlated with each other than with exon dispersions in other 
genes. Is it known whether this is the case or not, and is it possible 
to shrink exon dispersions towards two targets (mean gene dispersion and 
overall trend)?

-Ryan

On Mon Mar 31 22:16:23 2014, Alejandro Reyes wrote:
>
> Hi Ryan,
>
> Thanks for your e-mail!
>
> For all previous versions, the dispersion-mean trend was done using
> the mean for a given exon. The latest version on the svn is including
> the sum of counts from the other exons, but this is not optimal, since
> (as you mention) the mean of the counts get inflated and each exon
> from the same gene would have the same mean, for some cases, this
> causes the dispersion-mean trend not to be obvious. We will change it
> soon in the next commit of the svn, the reason we have not done it is
> because when doing it we encounter some problems with the running
> times in subsequent steps... my feeling is that it is a strange
> behavior of the GLM fitter, and I already contacted the maintainer,
> so it is work in progress!
>
> Will keep you updated,
> Alejandro Reyes
>
>
>>
>> Hello,
>>
>> I am trying to understand the methods used in DEXSeq, and I was
>> hoping you could clear up some of the details of the dispersion trend
>> squeezing strategy. In particular, when fitting the dispersion-mean
>> trend, what is used as the mean for a given exon? Is this computed
>> only based on the exon's own counts, or is it the mean of all the
>> counts for that exon, including the sum of counts in other exons? If
>> it is the latter, wouldn't the mean be identical for every exon in
>> the same gene, since the total counts for that gene are constant, and
>> if so, is this the intended effect?
>>
>> Thanks,
>>
>> -Ryan Thompson
>