[BioC] Using DESeq normalized gene count to replace FPKM?

[Simon Anders]

Hi Jack

On 23/07/13 17:29, Jike Cui wrote:
> A few papers have concluded that DESeq is more accurate for DE genes
> discovery than methods using FPKM, and that the bias in FPKM is that a
> gene’s FPKM depends on the expression of other genes due to the division by
> library size.
>
> Now if my purpose is for visualization or analysis other than looking for
> DEGs,  I wonder if it’s better to replace FPKM by DESeq normalized gene
> count divided by gene length?

Yes, definitely.

Look at it this way: To account for sequencing depth, you divide the raw 
counts by a number which quantifies this depth. Simply using the total 
number of reads (divided by 1 million) is an obvious but very simplistic 
choice, and the various other scaling normalization schemes (our 
median-of-ratios approach from DESeq, but also other similar suggestions 
such as TMM, etc.) are simply meant to suggest a more clever way to find 
a number to divide by.

In case of DESeq, we try to get this numbers to be close to one. If you 
want to have the same scale as typical FPKM values (and so have better 
comparability across experiments), you could then divide everything by 
something like

   geometric mean of the total read counts of all samples / 1 million


You may want to look, though, also at the variance-stabilizing 
transformation (VST) and the regularized log transformation (rlog) that 
we offer in DESeq2, and which, we feel, offers a better input for 
downstream visualization.

   Simon