Statistical comparison of more than two ChIP-seq experiments

[jwfoley]

I have a peaks × experiments matrix of tag counts from ChIP-seq experiments in several conditions (with replicates), and I want to know which peaks are differentially enriched in different conditions. Basically, for each row of the matrix, I want to do an ANOVA (or something equivalent) asking if the enrichment is the same in all classes or not. Of course, I need to transform the raw counts in some way first.

Does anyone know how to do this rigorously? I know DESeq, edgeR, and DEGseq are designed for this kind of analysis, but as far as I can tell they only support a two-class comparison. Is there a way I can use them to compare more than two classes? Would it be appropriate just to use their variance stabilizations etc. to transform my data, then replace their two-class test with an ANOVA?

Also, each of my experiments has its own total chromatin control, so each control is a separate column in my matrix. Can these programs use ChIP-seq controls in some way, or are they only really for RNA-seq?

[Simon Anders]

Hi

I have just added a new functionality to DESeq which allows you to fit generalized linear models (GLMs), in order to support more complex contrasts than simple comparisons. I am currently testing and documenting it and will release it soon.

Your application to ChIP-Seq was one of the use cases I had in mind for this GLM functionalioty but I could not try this out yet for lack of suitable data. As you have replicates and even per-sample input controls, your data might be a great test case. I'd be very interested to hear more details.

Cheers
Simon

[jwfoley]

Thanks. That sounds perfect. I'd be very interested in trying it as soon as it's available, even if it's not fully documented, since I have a meeting coming up and would like to have preliminary results to discuss.

I can give you an example of my data if that would help. Let me know how to send it to you.

[idonaldson]

I have a related question that i recently asked of the bioconductor forum. Should the ChIP-seq reads used for making the count tables be filtered to remove redundant reads (i.e. those mapping to exactly the same position)?

[jwfoley]

That question would be better off in a brand-new thread than a year-old one, but even better is to see whether anyone's asked it already. Here's one previous incarnation of it: http://seqanswers.com/forums/showthread.php?t=6543

[Boel]

jwfoley,
I have some data coming in which were produced in a very similar setting. How did you end up analyzing your data? Do you have any references for where similar comparisons have been made?
Thanks!

[jwfoley]

I developed a new method that is not yet publicly available, but I'm hoping to submit the manuscript within the next month or two and will make the method available once the paper is accepted. For the statistical analysis, I used DESeq.

[Boel]

I developed a new method that is not yet publicly available, but I'm hoping to submit the manuscript within the next month or two and will make the method available once the paper is accepted. For the statistical analysis, I used DESeq.

Haha.. OK, good for you. Hope things go smooth with the manuscript.