Denominator in RPKM calculation

[fbarreto]

Hi all,

While I understand there are many pros and cons of different normalization methods for RNA-seq gene expression data, I'm currently using RPKM as an initial measure of expression to compare between samples.

One issue that has crossed my mind is regarding the denominator in RPKM calculations, specifically the 'total mappable reads' (as defined in Mortazavi et al. 2008).
Does that refer to: (1) the total number of reads the mapping operation started with? or (2) the sum of reads that actually were mapped (ignoring reads that had no matches in the reference)?


For example, if I upload a total of 1 million reads for mapping, and only 500,000 were successfully mapped onto my reference (with the remaining being discarded), what is the denominator in RPKM, 1 million (option 1 above) or 500 thousand (option 2)?


My interpretation of many Materials and Methods descriptions, method (1) is used. However, it seems to me that method (2) is more appropriate, especially when using an incomplete transcriptome as reference (as is my case). What I mean is, the number of actually mapped reads (option 1) will depend on how many exons (or contigs) were available as reference. Hence, if the two samples being compared have different number of contigs in their respective references, their will different total number of successfully mapped reads even if their per-exon-mapping is identical. Their calculated RPKM will be very different depending on the denominator.

Any insight will be appreciated.

Thanks!

[Pseudonym]

One issue that has crossed my mind is regarding the denominator in RPKM calculations, specifically the 'total mappable reads' (as defined in Mortazavi et al. 2008).
Does that refer to: (1) the total number of reads the mapping operation started with? or (2) the sum of reads that actually were mapped (ignoring reads that had no matches in the reference)?

It is, as you suspect, the latter.

I should point out for completeness that in the case of de novo discovery, it's not necessarily obvious what a "mapped read" exactly is, since the "reference" is constructed from the reads. It's usually interpreted as the number of reads which "align" to the cleaned-up assembly structure (e.g. de Bruijn graph in the case of Oases, components in the case of Trinity).

My interpretation of many Materials and Methods descriptions, method (1) is used.

Do you have an example?