[BioC] DESeq2 1.2.10 vs 1.5.26

Michael Love michaelisaiahlove at gmail.com
Thu Jul 17 17:59:29 CEST 2014


hi Ido,

Nevermind, I was able to view the images on my cellphone. These
results look concordant to me, the p-values are very close to each
other and we have a bit more shrinkage of log fold changes due to some
recent improvements on the beta prior variance. Note that the
dispersion is simulated from a line disp = 0.1 + 4 * base-mean, so of
course the shrinkage towards the line (which the parametric curve has
accurately captured) makes sense.

best,

Mike

On Thu, Jul 17, 2014 at 11:35 AM, Michael Love
<michaelisaiahlove at gmail.com> wrote:
> hi Ido,
>
> I can't see these images for some reason. Maybe you can forward them
> as attachments to me?
>
> You jumped from the October 2013 release to the current devel version,
> and hence got a lot of improved estimates at once. Due to an
> improvement in dispersion estimation from 1.2 to 1.4, the maximum
> likelihood estimates in your dataset are now falling closer to the
> line (variance of log dispersions of 0.711 instead of 0.977), and
> hence the model "trusts" the fitted line more.
>
> In the development branch (so from 1.4 to 1.5), we have made more
> robust the estimate of the variance of the prior on log fold changes.
> Note that the development branch is just that: for development. So if
> you want more stability, you should use the release branch (v1.4).
>
> best,
>
> Mike
>
>
> On Thu, Jul 17, 2014 at 11:20 AM, Ido Tamir <tamir at imp.ac.at> wrote:
>> Hi,
>>
>> I tried to switch from DESeq2_1.2.10 (R3.0) to DESeq2_1.5.26 (R3.1)
>>
>> but it looks like it 1.5.26 is much more aggressive in the shrinking of the variance estimation.
>> The rlog normalized fold changes are also very different.
>>
>> The dataset was generated with
>>
>> dds <-  makeExampleDESeqDataSet(n = 30000, m = 6, betaSD = 1.5)
>>
>> saved and worked on in 2 different R/DESeq2 versions.
>>
>> a)
>> Now the obvious question is: is newer truthier?
>>
>> b)
>> Is there a parameter to get similar estimates with the new version as in the old version.
>> Some estimates are more robust now I read in the news e.g. Cooks distance, beta prior variance.
>> But I don’t understand the large changes this entails for some estimates.
>>
>>
>> thank you very much,
>> ido
>>
>> dispersion estimate plot 3.0
>> http://postimg.org/image/t3z7i27zz/
>>
>> dispersion estimate plot 3.1
>> http://postimg.org/image/v9tickbgf
>>
>> log fc 3.1 vs 3.0
>> http://postimg.org/image/mp06le1a7
>>
>> mean 3.1 vs 3.0 illustrating identical input data (< 100)
>> http://postimg.org/image/hno76a4fz
>>
>> mean 3.1 vs 3.0 illustrating identical input data (all)
>> http://postimg.org/image/wv46qmwan
>>
>> pvalue 3.1 vs 3.0
>> http://postimg.org/image/7rn46mynz
>>
>> 3.0:
>> attr(,"coefficients")
>> asymptDisp  extraPois
>>  0.1147892  5.3433952
>> attr(,"fitType")
>> [1] "parametric"
>> attr(,"varLogDispEsts")
>> [1] 0.9778035
>> attr(,"expVarLogDisp")
>> [1] 0.6449341
>> attr(,"dispPriorVar")
>> [1] 0.3328694
>>
>> 3.1:
>> attr(,"coefficients")
>> asymptDisp  extraPois
>>  0.1117856  5.8477354
>> attr(,"fitType")
>> [1] "parametric"
>> attr(,"varLogDispEsts")
>> [1] 0.7119179
>> attr(,"expVarLogDisp")
>> [1] 0.6449341
>> attr(,"dispPriorVar")
>> [1] 0.25
>>
>> _______________________________________________
>> Bioconductor mailing list
>> Bioconductor at r-project.org
>> https://stat.ethz.ch/mailman/listinfo/bioconductor
>> Search the archives: http://news.gmane.org/gmane.science.biology.informatics.conductor



More information about the Bioconductor mailing list