Calculating RPKM value manually

[ThePresident]

@Ryan: I totally agree with you and that was my fear before I run alignments with bowtie2. However, i reasoned that since I'm dealing with bacterial transcriptome, the complexity of the reference genome is way smaller then that of human RNA-seq. So, I didn't expect a huge number of multimappers as defined by the --score-min function (i.e. reads that have more then one valid alignment).

I turns out that I was probably right: I lose about 3% of my total reads by filtering for multimappers with XS: tag which is not dramatic. However, one library was more of a mess... It had a deepest sequencing depth (61 M reads) but I lose 14% on "multimappers".

In my opinion, bowtie1 had somewhat more flexible options. Reporting could be controlled with "-a --best --strata" or "--best -k 1".

TP

[dpryan]

@ThePresident: That makes more sense then. Yeah, bowtie1 did have its advantages.

[bswhite]

I calculated RPKM values manually, still didn't compare it with Cufflinks RPKMs, but visually I have a lognormal distribution. Sounds logical for me...

TP

If/when you do compare to cufflinks RPKM, I would be interested to hear
the results.

I also see a (skewed) lognormal computing RPKMs via htseq, but I
see a bimodal lognormal from cufflinks FPKMs. [Yes, I mean RPKM
in the first and FPKM in the latter--an oversimplification on my part
using these paired-end data.]

[raymb]

Can someone please explain where the "10^9" in the calculation comes from? Why isn't it 10^6?

Thanks in advance!

[dpryan]

Can someone please explain where the "10^9" in the calculation comes from? Why isn't it 10^6?

Thanks in advance!

L is usually measured in base pairs but the RPKM/FPKM metric requires things in KB, so the 10^9 comes from that conversion. If you input a length measure in KB then just use 10^6.

[jhb1980]

I'm curious about calculating RPKM / FPKM from counts on the gene level. What would you use as "gene length" for a gene with multiple isoforms...? The longest exon chain (i.e. longest transcript isoform)? Or a median transcript length based on all the annotated isoforms of a given gene...?

[dpryan]

Usually one uses the union of all exons and gets the length of that. At least unless you want to determine an expected gene length given alignment data (some tools can do that).

[wanfahmi]

Can we get the FPKM value only without through a long process of Cuffdiff? I do have 100 samples and each sample has 3 biological replicates. I am only interested in FPKM value which will later use for network analysis. Thanks

[dpryan]

Yes, but there's more than one way to get and even define FPKMs. You could just get the counts with something like featureCounts and then convert them to FPKM (either before or after normalization). That's the fastest method. You could also get expected counts with something like eXpress and convert those to FPKMs. That would take longer and given you different (through pretty similar) results. I can conceive of still other routes.