[Simon Anders]

In an RNA-Seq context, this is no problem at all; DESeq and edgeR are designed to handle this implicitly. I'm not sure what you mean by oversampling, but I guess it is nothing that happens in RNA-Seq. There, you may get pretty uneven coverage across a gene, but if you look at tracks for two replicates, even if they have different depth, everything scales nicely and linearly (except for the shot noise, of course, which scales with the square root).

[mudshark]

Hi Simon. RNASeq does/might not suffer from oversampling but ChIPSeq does. Oversampling appears to happen in particular in regions of high DNAse sensitivity, i.e. promoters and regulatory regions. That's probably due to a preferential release of sequencer-preferred-size fragments there. While transcribed sequences are under-sampled. All tested using qPCR.

[biznatch]

For input, SISSRS considers our input sample before calling peaks so any "false" peaks shouldn't even be called. Though as Simon said, when looking for changes between treatments, false peaks shouldn't be changed so it probably isn't much of an issue.

In regards to normalising between samples (treated and untreated), in this case, they have nearly identical number of reads and number of peaks so it's not an issue, we also plan to validate sites by ChIP-PCR. But it is something I've thought of for future experiments, I've tried looking online but haven't found any consensus on what to do.

What if normalisation is done for each peak comparison individually? Ie. look at the 500bp up and downstream of the peak, determine average height, compare between samples and scale one of the samples linearly, as needed? Would this account for both changes in overall sequencing depth and variability in particular regions of the genome?

[AdamB]

@Simon Anders
Would DESeq apply linear scaling between treatments (rather than between sample X and input X)? Please can you elaborate more on what you would like to see in a scatter plot? Also, are replicates required or can two individual samples be compared?

@biznatch
If I plot the number of aligned reads at TSS ±2 kb, my treatments look similar, but there is a definite difference across the region - one treatment is higher. I'm unsure whether this is a real global difference in H3ac (its possible my treatment could induce such an effect), or due to biological/technical variation. I guess the only way to determine this for sure would be with a Western blot?

[Simon Anders]

Adam: Each column of the count table gets a normalization factor. The table has one column for each ChIP sample. The input sample does not enter the count table. This is what I meant by saying that the input should only be used for finding the peaks, not for comparing them.

[biznatch]

Adam, just to be sure I understand, you find the same number of reads between samples at TSS ±2 kb, what region are you referring to that you see a definite difference? Do you mean that in the local region around the ±2 kb, like, within 100kb, you see a difference, or do you mean one sample is increased across the entire genome?

[AdamB]

I mean there is an enrichment of reads at the regions surrounding the TSS, relative to the rest of the genome. One sample isn't consistently increased across the whole genome, since I'm normalising to total reads (to account for sequencing depth) for this comparison.

If I count the reads across regions at least 10 kb away from genes (i.e. to find regions of predominantly background) the sample that has more reads surrounding the TSS has less reads in these regions. This suggests there is higher signal:background in this sample, due to either a real biological difference, or difference in ChIP enrichment levels.

[biznatch]

Ah ok I understand. So if there is a real biological increase in your histone at/near the TSS then I think you could be expected to see a decrease in the background. Eg. if you do 10 million reads of each sample, and one sample has lots of peaks at TSS then the TSS reads will take away from the background reads.

I guess this is why it's hard to quantitatively compare samples that have large differences in peaks. The sample with lots more peaks will appear to have less background so normalising to (local) background levels may not be accurate. Maybe come combination of normalising to total number of reads and local background...have to keep thinking about it...

In your case I'd say a western is probably a good way to determine if it's a real biological increase. And/or check a few sites by ChIP-qPCR.

[hongbo919]

A professional tool for identification of differential histone modifications

Recently, we developed a tool EpiDiff including a professional tool QDCMR for identification of differential histone modifications.
The software of EpiDiff is available at http://bioinfo.hrbmu.edu.cn/epidiff/. Look forward to you for testing this software and helping us to improve the software.

[mudshark]

Recently, we developed a tool EpiDiff including a professional tool QDCMR for identification of differential histone modifications.
The software of EpiDiff is available at http://bioinfo.hrbmu.edu.cn/epidiff/. Look forward to you for testing this software and helping us to improve the software.

hi! unfortunately your server seems incredibly slow. i did not manage to download the standalone version. in addition, it would be nice to get some information on how your software works. do you have this documentation online, or published?

[hongbo919]

hi! unfortunately your server seems incredibly slow. i did not manage to download the standalone version. in addition, it would be nice to get some information on how your software works. do you have this documentation online, or published?

Hi. Thanks for your test. You can try http://202.97.205.78/epidiff/ for the EpiDiff. In addtion, the paper of this tool is under review and has not been published. However, there are detailed tutorials which can be found easily in the website . The turorials will tell you how this software works.
Best wishes!

[asiangg]

UPDATE: We have developed diffReps to detect differential chromatin modification sites from two groups of comparing ChIP-seq data.

https://code.google.com/p/diffreps/

It takes into account of biological replicates. It bins the genome, calculates counts, normalizes by linear scaling, performs statistical tests and adjusts p-values. It does all of these in one command line. I hope people may find it to be useful!

I'm not sure if we'll publish it in a paper. But I just want to share it with the community as an open source software.

[EdinG]

Detecting Shape Changes in ChIP-Seq data

Hi,

we have just added a new package called MMDiff to the latest Bioconductor release.
It's a statistical testing method specifically designed to compare ChIP-Seq data sets. It takes advantage of higher order features to detect shape changes in ChIP-Seq peaks and should also be applicable to other -Seq methods, like DNase-Seq. Biological replicates are used to estimate biological variance.

Here is the link to the package:

http://bioconductor.org/packages/release/bioc/html/MMDiff.html