Statistical comparison of chip-seq for 2 conditions

[epi]

I have 2 conditions, under which I want to find histone modifications that are different. I have 4 samples in total, Condition1 and its control, and condition2 with its control.

SICER does it just fine.

I was wondering if there is any other tool anybody knows of. In fact, It is easy to find peaks using for a condition and its control (or background distribution), is there something available that can take already identified peaks (bed, GFF whatever) and reports differential reads counts with statistical significance. I looked at few, but nothing that works to my satisfaction:

ChIPDiff Apparently does exactly what I am looking for, however uses only single genomic position as against chr, start and end. Seems apt for TF but not histones because of variable lengths. In addition asks for orientation which my peak finder does not report. Just not a good fit for my data I guess.

DESeq finds differential expression using counts, but relies on multiple samples to report significance. And needs gene-expression like matrix which I will have to figure out how to get from the peak data.


I know this has been already asked here, and I apologize for creating a new thread. But it seems the discussion is still to reach climax (as if it can ever ) and new tools and papers are coming every day so hoping there is something by new out there.

Thanks for replying.

[ETHANol]

This is what I do. It's not statistical but it seems to find differentially bound regions pretty well.

http://ethanomics.wordpress.com/2011/08/07/how-to-compare-transcription-factor-binding-between-two-chip-seq-samples/

[epi]

Thanks for the replying ethanol. Good to hear your ideas.

It seems you are suggesting to (1) first do a peak finding cond1 Vs input, (2) then cond2 Vs input, and then peaks found in step 2 vs 1.

This is a clever idea.

However, I have some doubts, depending on which peak finder you are working with. Nevertheless, it can be one useful way to look at the data.
Diffbind seems interesting, will check it out.

[mudshark]

maybe you also find useful information in here

http://www.nature.com/nprot/journal/v7/n1/full/nprot.2011.420.html

you should also consider replication of your experiments! two sample comparisons without replication are be pretty shaky if your effects are not bloody obvious and cannot easily be verified using qPCR

[ETHANol]

maybe you also find useful information in here

http://www.nature.com/nprot/journal/v7/n1/full/nprot.2011.420.html

Have you looked at that publication? Buggy, some platform specific code, sample name specific code, some formatting mistakes, hard to understand documentation, the novelty is questionable....
Some more work and the probably could have made some generally useful tools.

[mudshark]

@*ethanol: haha, right, you got me. did not look at the publication, yet.
@ epi: sorry epi.

simply thought that a nature protocols protocol would be reviewed thoroughly. naive thinking..

[Dario1984]

Have you looked at that publication? Buggy, some platform specific code, sample name specific code, some formatting mistakes, hard to understand documentation, the novelty is questionable....
Some more work and the probably could have made some generally useful tools.

Sounds like the recent RNASEQR pipeline. They hard-coded the chromosome names, so it produced errors for anything other than human samples. The names are also in ENSEMBL format, so a RefSeq annotation also causes problems.

[kopi-o]

The ones I know about are

- ChIPDiff, which you already mentioned but which I haven't tried, but it was actually developed for histone marks so I'd be surprised if it didn't work for those?! I recall someone saying that there was some other issue with it (sorry I can't be more specific)

- DiffBind, which you also mentioned (http://www.bioconductor.org/packages...iffBind.html); it uses DESeq internally

- DBChIP (http://pages.cs.wisc.edu/~kliang/DBChIP/), which appears to use edgeR

- You can also use edgeR and DESeq directly. The DESeq paper shows you how to re-analyze differential TF binding data in the Kasowski et al Science paper (http://www.sciencemag.org/content/328/5975/232.short).

- That Kasowski et al paper in itself shows a GLM (generalized linear model) based method to do such a comparison, check out the Supplementary Methods

[ETHANol]

Sounds like the recent RNASEQR pipeline. They hard-coded the chromosome names, so it produced errors for anything other than human samples. The names are also in ENSEMBL format, so a RefSeq annotation also causes problems.

Seems everyone else like myself is on the wrong side of the Natrure Protocols paywall. Annoying. Anyway, this pub is like some random code I typed in while trying to figure out a good ChIP-seq analysis pipeline and makes RNASEQR look like a highly refined highly flexible piece of software. I'm not sure why the call RNASEQR an analysis program when all it does is map the reads.

[epi]

Seems everyone else like myself is on the wrong side of the Natrure Protocols paywall. Annoying. Anyway, this pub is like some random code I typed in while trying to figure out a good ChIP-seq analysis pipeline and makes RNASEQR look like a highly refined highly flexible piece of software. I'm not sure why the call RNASEQR an analysis program when all it does is map the reads.

I have not tried it, but i understand what you mean. I have similar experiences with some R/bioconductor packages (published as journal articles). Minimal documentation and hard coded stuff to run well with example data, but very difficult on real data.

meanwhile, I am working with diffbind, will share my experience once I am through with the analysis.

[epi]

- ChIPDiff, which you already mentioned but which I haven't tried, but it was actually developed for histone marks so I'd be surprised if it didn't work for those?! I recall someone saying that there was some other issue with it (sorry I can't be more specific)

Thanks Kopi-o for your additions.

My concern is that ChIPDiff asks for a single genomic position as the peak identifier. Now many histone modifications are variable length (long and short range, in fact some can be very long). Its scary to try it considering I will loose all information about my peak length.

Something else surprised me, that ChIPDiff is developed by the same team who did CCAT peak finder. Now if you run CCAT for 2 different ChIP-Seq and want to find quantitative differences across them, you can't because CCAT does not give you back the orientation (+ or -), which ChIPDiff needs.

I hope I am not missing something obvious about the usage, in which case I will appreciate clarification.

[kopi-o]

It's not only the same team, but the same person, who wrote those two progs by himself (more or less). I think you should simply email him (the first author on the ChIPDiff paper) and ask about the usage.

[epi]

diffbind works well.
Thanks for suggesting it, and will be happy to discuss if anyone wants to use it ...