Normalizing with ERCC spike-in

[Eric Fournier]

Hello everyone,

I have RNA-seq libraries prepared from 10 different stages of embryonic development (3 replicates per stage), with each library constructed using the same number of embryos. The amount of total RNA should be variable between stages and follow a known pattern. Thus, to be able to compare "absolute quantities" of RNAs, we spiked each of the libraries with ERCC controls after RNA extraction, but prior to any other processing step.

I've aligned the sequencing results using TopHat2, and constructed transcripts using cufflinks. However, now that I'm getting to the "normalize using ERCC" steps, I am a unsure on how to proceed.

My first instinct would be to do a regression of the ERCC's FPKMs against their known concentration for each of the libraries, then report all of the other transcripts' FPKM against that curve. However, given that FPKMs are already a normalized value, is this still a good idea?

Furthermore, going from FPKM to whichever measure I obtain will make it impossible to use standard RNA-seq comparison tools, such as cuffdiff. Would there be another kind of normalization which would be more "standard" or more sensible?

Thank you for any help,
-Eric Fournier

[jparsons]

The easy route is to just follow the "method" used in the Cell paper (Revisiting Global Gene Expression Analysis, 151, Oct 2012). They do more or less exactly what your first instinct suggests, fitting a regression of the ERCC spikes and renormalizing.

When running any cufflinks/cuffdiff analysis in a sample which contains ERCCs, you don't want to keep ERCC-mapped reads in the denominator of your FPKM calculation. You could either normalize them away (by multiplying through by total reads / Total non-ERCC Reads) or you could prevent them from showing up in the first place (by mapping them separately) and then factoring in their relative ratios after the fact. I prefer the latter method, because I don't understand everything that Cufflinks does in its calculations, and I don't trust that the presence of spiked-in RNA doesn't cause one of Cufflinks' calculations to make an assumption that isn't true in my sample.

The renormalized values would still be FPKM, as you are merely correcting for the incorrect assumption that Cufflinks initially makes about your sample. You should be able to carry forward with Cuffdiff after you change the denominator to the proper value.

[Eric Fournier]

Thank you very much! The article was a very nice read.

[danwiththeplan]

Eric I'm curious, did you multiplex all your samples and run them on a single lane? So you required 30 separate spikes, for each library, which were then multiplex-tagged and combined? Or were they all on a different lane?

Also, did you use the ExFold mix to look at fold-change effects?

[Eric Fournier]

Hello Dan,

we ran our samples on five different lanes. Each lane used 6 of 8 possible multiplex tags from the Encore multiplex kit (which uses 4nt tags). This actually caused a small problem, since one of the combination of 6 tags that we used caused library complexity for the first four nucleotides to go down substantially, which was reflected as low quality values across the whole library.

The ERCC spikes were added immediatly after RNA extraction, while the multiplexing was done just prior to sending the libraries to the sequencing center.

Since we had 10 different tissues and that we were not interested in any particular pairwise comparison, we did not use th Exfold mix to assess fold-change effects. Rather, we used only mix 1 from the ERCC to have one shared standard across all libraries.