Peak height for GAPDH in mouse

[dnusol]

Hi,

has anyone looked at the peaks detected for GAPDH after a histone ChIP experiment in mouse? what height do you usually get? Is it a good control for filtering other peaks?

Cheers

[simonandrews]

GAPDH is a pretty awful control for anything really. There are loads of partial copies of the gene throughout the genome which makes assigning reads to a specific copy very difficult:

http://www.ensembl.org/Mus_musculus/Gene/Compara_Paralog?g=ENSMUSG00000057666

The reason it used to work well as a control in array based experiments was because you were actually measuring signal from all of the different copies so you got a kind of global normalisation in a single probe. Once we moved to sequencing based methods it stopped being such a useful control.

[vruotti]

Hi Simon,
We see some variation when doing RNA Seq on human. We read your comment. However, if we look don't allocate reads to individual isoforms then I don't understand how this is different from the global normalization in a single probe.
We used our EM algorithm to allocate reads from our RNA Seq experiment to genes.

The interesting thing is that we do see some variation in the expression (TPMs) of GAPDH across samples. We think this is due to the higher accuracy of RNASeq compared to Microarray or even QPCR.

What do you think?
Victor

[simonandrews]

Hi Simon,
We see some variation when doing RNA Seq on human. We read your comment. However, if we look don't allocate reads to individual isoforms then I don't understand how this is different from the global normalization in a single probe.

The problem was that array based hybridisation methods actually picked up signal from more than one of the isoforms, which is partially why GAPDH was so stable. The sensitivity of sequencing means that you can now discriminate between isoforms, and that individual forms may not be as stable as array methods suggested. In your case GAPDH may be a very good control but we've seen groups which use it just because it's a name they know without realising why it was so good originally. Incidentally the same is true (to a lesser extent) for Actin.

The interesting thing is that we do see some variation in the expression (TPMs) of GAPDH across samples. We think this is due to the higher accuracy of RNASeq compared to Microarray or even QPCR.

Firstly I think it's a really good thing that you're looking into this. We're increasingly finding that variations in ChIP efficiency between different samples are potentially giving us false differences, or masking real differences in our data. Our normal approach is to globally normalise the data so that the median counts over peaks are consistent across samples. However this assumes that the overall level of signal between samples is the same - which is probably true in our case, but may not be in yours. Your plan to use a marker of known enrichment to normalise is therefore a very good alternative.

I suppose the problem of variation boils down to how sure you can be that GAPDH is really stable in your system. Have you compared the GAPDH variations to the global median variation and see if the two agree? I suppose we could go back to the ideas behind the GNorm method which we use for RTPCR assays. Basically you can provide this with a set of enrichment values over multiple samples and a list of potential control genes and it will find the most stable subset of those genes which can then be used to normalise the rest of the data.