Normalisation of chIP-seq data for quantitative comparison

[albireo]

Hi Rory

thanks a lot for the great info.

[rory]

This is a very good question, and while it does not yet have a definitive answer, there is literature suggesting that differentially bound sites have biological significance.

For example, we have been able to a) identify differentially bound sites based on statistically significant differences in peak density between replicated sample groups; b) show systematic changes in gene expression near differentially bound sites; and c) demonstrate predictive power for these expression profiles in determining phenotype (specifically disease outcome). See Ross-Innes et al "Differential oestrogen receptor binding is associated with clinical outcome in breast cancer." Nature 2012

[albireo]

I agree with your points.

My concerns have more to do with tthe idea of doing quantitative chip-seq differential analysis. Even supposing we do have replicates, isn't chip-seq considered semi-quantitative at best? E.g. is there sound evidence implicating peak-size to biological effect across cells/populations/conditions? I am rather new to the field and I might have missed relevant literature - apologies if that's the case.

[rory]

Here's my understanding of why we need replicates:

There is substantial variance in read density at specific peak sites between biological and technical replicates. Without replicating the experiment, the variance can not be measured (or even estimated accurately). As the confidence statistics (p-values, FDR) computed by the differential analysis techniques are dependant on accurate measurements of variance, they can not produce reliable results in the absence of replication.

Interestingly, there is very little questioning of the need for replication in differential RNA expression analysis, whether it be via microarray or RNA-seq, but the question comes up all the time in ChIP-seq, where the data demonstrate even higher levels of between-replicate variance.

[albireo]

For the original poster, without replicates it is not possible to perform a meaningful quantitative differential analysis.

Hi Rory, would you care to explain why it is not possible?

[rory]

The Bioconductor package DiffBind has a number of normalization options, with the default using the edgeR TMM (based on effective library size after subtracting control reads). It makes it easy to do some boxplots to see the effect of the normalization. DiffBind will also run single- and multi-factor differential analysis using both edgeR and DESeq.

For the original poster, without replicates it is not possible to perform a meaningful quantitative differential analysis.

[albireo]

Ok thanks, I had missed that feature of Chance. So would you say DESeq could work on Chip-seq data? I thought it was tailored to RNA-seq.

[kopi-o]

CHANCE has a multi-IP normalization mode, see Figure 3 in the paper: http://genomebiology.com/2012/13/10/R98

Aside from the aforementioned MANorm and just using DESeq's or edgeR's in-built normalizations, there is also SeqMiner which performs normalization.

[albireo]

Hi, does anybody have any suggestions about this? I'm facing a similar problem as the OP's and there seems to be a lack of tools for normalisation of chip-seq data

Also, I'm not sure the CHANCE suggestion is fitting - from what I've seen CHANCE mainly tells if the ChIP experiment was successful, providing figures for the S/N ratio between IP and control.

[leeloo047]

Yes I've tried DESeq but as I've no replicate I don't know if I can rely on these results which show very few significant differences between my samples... as if the differences were reduced by normalization.

[jwfoley]

Have you tried DESeq? You'll have to use raw read counts rather than peak heights.

[leeloo047]

Unfortunately, I'm working on S.cerevisiae genome and CHANCE only works with reads mapped to mm9, hg18, hg19, and tair10... So I can't use it...

[leeloo047]

I've already tried MAnorm but it is not appropriate for my data because the normalization is based on common peaks, however most of my peaks are common because it's always the same strain, with only some mutations which affect peaks height, and in some case make them disapear but it's not the majority of them.
I don't know the CHANCE package, I'm gonna try it, thanks.

[kopi-o]

Have you tried MANorm (http://genomebiology.com/2012/13/3/R16/abstract)? You might also want to look into the CHANCE ChIP-seq quality control package (http://www.ncbi.nlm.nih.gov/pubmed/23068444).

[leeloo047]

Hi,
I would like to know if there is a way to normalise ChIP-seq data for quantitative comparison between peaks height. I've read a lot of paper on different methods but none seems to fit my data.
In fact, I have one WT strain and 3 mutant strains for which we can expect a global defect in transcription (all peaks decreased) or gene specific defect (only some peaks decreased). In all cases, we don't have any reference gene for which we know that we will not have any effect. Moreover, relying on the total number of reads could result in a reduction of the differences between WT and mutant strains in the case of a global effect.
Anyway I seem to have almost tried everything without success and I'm running out of ideas... so if someone could help me.
Thanks.