How to compute RPKM?

[liuxq]

Everyone knows the formula for RPKM compuation: rpkm=10^9*C/NL，where C is the reads number of the transcript, L is the length of the transcript and N is the total reads number of the sample

However, in my RNA-seq analysis pipeline, I have three "N".

1. total reads number
2. number of reads which can be mapped to reference genome
3. number of reads which are the result after mappable reads filtering using repeatmask

how to select the total reads number N for RPKM computation? I find that using three "N" have totally different effect.

Thanks very much.

[RockChalkJayhawk]

Everyone knows the formula for RPKM compuation: rpkm=10^9*C/NL，where C is the reads number of the transcript, L is the length of the transcript and N is the total reads number of the sample

However, in my RNA-seq analysis pipeline, I have three "N".

1. total reads number
2. number of reads which can be mapped to reference genome
3. number of reads which are the result after mappable reads filtering using repeatmask

how to select the total reads number N for RPKM computation? I find that using three "N" have totally different effect.

Thanks very much.

If all your experiments use repeat mask, then use option 3. Just make sure to clearly point out this definition when you report FPKM.

[liuxq]

If all your experiments use repeat mask, then use option 3. Just make sure to clearly point out this definition when you report FPKM.

why using option 3 is more reasonable?

[RockChalkJayhawk]

Option 3 in this scenario represents the last step in processing - or the final number of mapped reads that you will use in your analysis. It is not as informative to use any N other than what passes through your quality control steps.

[sameet]

Hi,
I am a bit confused. What should i use for N, total number of reads that mapped, or the unique number of reads that mapped. I cannot afford to discard the repeated reads because I have some important data in it.

[RockChalkJayhawk]

Hi,
I am a bit confused. What should i use for N, total number of reads that mapped, or the unique number of reads that mapped. I cannot afford to discard the repeated reads because I have some important data in it.

In that case I would use 2, but make sure you clearly state that you haven't removed reads from repeat regions.

[sameet]

In that case I would use 2, but make sure you clearly state that you haven't removed reads from repeat regions.

Hi,
I was thinking along same lines. But I want to know how to handle situations when the same read maps to multiple locations, because this happens at a a pretty high high rate in my samples.

[severin]

Hi Sameet,
As far as I have seen there really is no clear rule on what to do with mappings to multiple locations, which is why many scientists use uniquely mappable reads for each gene. In the RNA-Seq Atlas for Glycine max, I used the uniquely mappable reads then use the mappable total count (N) that includes the multiple alignments. Now ,of course, there are programs (Cufflinks or Erange) that try to account for multiple mappings but that doesn't help you decide to include them in the first place.

As people above have mentioned, reporting the methodology is very important. I found in soybean (it has had two whole genome duplications, so lots of similar genes) the Atlas paper using only the uniquely mappable reads on a non-replicated sample still provided plenty of interesting data that fit what we would expect from a soybean (genes involved in seed filling still were highly expressed in seed filling etc).

No one method is going to be better than another in every case. It really depends on what you are looking at. Just be aware of the potential biases and include those in your interpretation.

[Simon Anders]

See Robinson and Oshlack's paper (Genome Biol 2010, 11:R25) for some thought why neither of the three 'N' values may be a good option, at least if you want to see differential expression.

[john23]

See Robinson and Oshlack's paper (Genome Biol 2010, 11:R25) for some thought why neither of the three 'N' values may be a good option, at least if you want to see differential expression.

RPKM/FPKM is a better option then the raw read counts because it takes into account the quantity of RNA which has been used for sequencing. In general the RNA samples are sequenced using different amounts of RNA which gives totally different number of reads (a larger quantity of RNA gives a larger number of reads).