[BioC] DESeq: nbinomTest question

[Simon Anders]

Hi Natasha

On 08/30/2011 11:48 AM, Natasha Sahgal wrote:
> I may be wrong here, but I wanted to know if the output of DE calling
> would differ if the normalised count were to be used rather than the
> raw counts as is the current case (using the cds - countDataSet)?

The CountDataSet object must be constructed with raw, unnormalized 
counts. The test takes into account the normalization, of course -- 
namely by accounting for the size factors. (See the papers for 
mathematical details.)

If you want to use an alternative method of normalization, you can do 
so, provided your method yields a single scaling factor for each sample. 
Just put these scaling factors in the sizeFactors slot instead of 
calling 'estimateSizeFactors'.

> As in the vignette, when trying to plot the heatmap, you rather use
> the vsd object (as you mention it is better for visual purposes). I
> was wondering if it would affect the DE calling? Or perhaps it does
> not matter because the baseMeans are calculated by dividing by the
> sizeFactors, as in a way it is normalised?

The variance stabilized values are not a proper input for a negative 
binomial test. They are meant to be used in any kind of statistical 
method that requires homoscedastic values, as does, for example, the 
'heatmap' function. You may use them in an ordinary t test but this 
would not be very useful, as the NB test gives better results.

   Simon