TSS plot. Drop of the signal at TSS.

[neurongs]

Hi all,

I'm fairly new to NGS data analysis. My question is related to the shape of the coverage plots relative to TSS of ChIP against an histone modif (i.e. histone acetylation). I have several datasets of an histone modif showing the "clasical" pattern observed in many papers (i.e. Barsky et al., 2007). That's a net enrichment of the signal at TSS regions. In these plots there is a sharp drop of the signal exactly at TSS boundary. This initial set of samples were processed with GAII technology. I always though that the cause of such a drop at TSS was because there is a number of TSS in which there is not histones at that position and this was also the reason by which the input usually shows an "enrichment" at TSS regions.

Some time latter I got additional datasets that were performed against the same histone modif, with the same antibody and the same mouse line but with different technology (HiSeq 2000). The point is that although the global profile of the samples is conserved, the coverage plot around the TSS is strikingly different. In both groups of datasets there is a nice enrichment of the signal around TSS genomic regions but the sharp drop of the signal at TSS position is gone in the second set of samples and the shape of the plot is therefore clearly different...

Both groups of data sets have a similar depth of sequencing (7-8 mio reads). I'm not sure but I susppect that the library construction methodology may have some influence... does anyone have an idea of what can be the reason? any hint would be very appreciated.

[mudshark]

A) did you also sequence the input? how do enrichment profiles look like (IP over input)? did barsky sequence the input?

B) differences in shearing also can account for substantial differences. your conditions/machines differ?

[neurongs]

Thank you for your answer mudshark,

Yes, I also sequenced the input and I also see enrichment at TSS as shown i.e. in http://www.ncbi.nlm.nih.gov/pubmed/19367334 but, in this case, there is not the sharp drop of the signal at TSS boundary and the plot has the shape of a unimodal distribution.

The point of my question is that I have TSS plots from a chip against a histone modif having a bimodal distribution shape (tech was GAII) (two biological replicates, both with the same pattern). Later on, I decided to resequence the samples again. New libraries were constructed and sequenced on a HiSeq machine. What I got is that the TSS plot of the new dataset does not have this bimodal shape but a unimodal one.... (in both cases there was an increase of the signal at TSS but with different shape).

I don’t fully understand what is going on but though that I am not the first one making this observation. I have discarded the involvement of the chip procedure (the sample to make the library is the same), the depth of sequencing (similar read number in the dataset), alignment differences (same tool (BWA) and filters (RSamtools) and soft-related concerns (both were done with the same tool (i.e. R (HTseq)). So, I think it may be due to advances in library construction methodology or sequencing chemistry (i.e. dealing with GC rich sequences quality...) but I am a bit as a loss as to how to interpret the differences between the TSS plots from the different technical replicates.

Again, thanks for your kind advice,

[mudshark]

if you look at individual instances (promoters/TSS) are these differences also apparent?

can you post the images of your alignments?

[neurongs]

In red, the original seq (GAII, 2008) in blue, the reseq doing a new library of the same sample (HiSeq, 2011). In grey, the input (HiSeq, 2011).

http://dl.dropbox.com/u/14906265/plot.pdf

[frozenlyse]

It looks to me that your ChIP enrichment worked a lot better in your HiSeq sample - the signal to noise is a lot better.

[mudshark]

a) the number of mapped reads and the read length was the same?

b) did you check on individual promoter instances? in your plot you look at a mean that is quite likely picking up some outliers. maybe your sequencing depth is not enough and the resulting noise is the reason for the differences. can you actually plot the mean or the confidence interval in addition to the mean? that should give you some indication.

[neurongs]

As you know, I am doing TSS plots of identical chip samples (biological replicates) performed against a histone modif. I sequenced first in 2008 with GAII technology and then I resequenced the same samples (but did new libraries) with HiSeq2000 technology. What I observed is that there was a drop of the signal around TSS boundary only in GAII-derived datasets.
Now, I have performed an independent analysis at CpG islands and found that there is a drop in the signal in the GAII datasets and not at the HiSeq machine. I think that the drop may be well explained by an increase in the sequence quality over time and that CG rich fragments were excluded by the restrictions imposed to the mapping algoritm (BWA)… HTH to other ngs members.

http://dl.dropbox.com/u/14906265/TSSplot%26CpGisland.jpg