[BioC] DESeq vs DEXSeq

[Simon Anders]

Dear Robert

On 03/26/2012 04:26 PM, Robert M. Flight wrote:
>> From my understanding of the two packages, DESeq (and alternatively
> edgeR) allow testing for diff. expression of any object one can define
> counts for, whereas DEXSeq looks for genes (however defined) where
> there are only one or a few exons that show differential expression.

The crucial difference between DESeq and DEXSeq is that the latter aims 
to tease apart changes to the overall expression strength of a gene and 
changes to only some of its exons. Conceptionally, we consider the for 
each sample the fraction "number of reads overlapping with the exon (or: 
counting bin) under consideration" over "number of reads mapping to any 
exon of the gene". If the gene's overall expression changes but the 
relative abundances of the different transcripts stay the same, these 
fractions do not change, and DEXSeq will not call this counting bin 
significant even if its absolute count does change significantly.

(Note that this is a simplified explanation of what DEXSeq does 
conceptually. To see what it actually does, please see our preprint on 
Nature Precedings.)

> My initial belief was that DEXSeq was the best choice, however we are
> working with data from Rat, which has rather poorly annotated exons,
> especially in non-coding regions (i.e. UTRs). Therefore, I am thinking
> of defining exons based on a combination of the current annotation,
> known UTRs, and exons assembled by CuffLinks. I am not sure how this
> set of exons would fit into DEXSeq, and it seems to me that DESeq
> would be more appropriate, with determination after DE analysis to
> determine exon location (CDS, UTR, etc).

Once you have defined exons on a combination of information you trust, 
you can use DEXSeq. All you need is a table of counts, one column for 
each sample and one row for each exon -- or for whatever counting bins 
you want to define: It may be useful, for example, to keep the UTR and 
the coding part of outer exons separate. Then, define a factor to 
indicate which rows belong to the same gene and use this to call 
'createExonCountSet'.

   Simon