450K - "Best" Normalisation Method?

Question

[andrew.j.skelton73]

Hi,

I've got some 450K data and I've been looking at the various packages around....There's a lot of normalisation methods.

I realise normalisation as a method is quite subjective to the data, and it's probably best to "shop around", and see which method best fits my data. However, are there any methods that in principle should be better than others?

Thanks,

Answer 1

[fortin946]

Hi Andrew,

you are corrected, depending on your dataset characteristics, some normalization methods are more appropriate than others. If the goal of your analysis is to compare different tissues, cell types, or cancers versus normals (or any other dataset with global methylation differences), you might want to take a look at our recent paper comparing normalization methods for such datasets: http://www.genomebiology.com/2014/15/12/503. We showed that Functional normalization performs really well at conserving global methylation differences (preprocessFunnorm() function in the minfi package)

Otherwise, if you are doing a EWAS-type of study, you may want to use preprocessNoob() + preprocessQuantile(), both implemented in the minfi packages as well. The former performs a background correction based on the negative control probes (implementation of the noob method described in http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3627582/) that is always appropriate to use (and btw is called internally by preprocessFunnorm as well). The preprocessQuantile() function is an adaptation of quantile normalization to the 450k array, very appropriate when no global differences are expected between your samples (for instance whole-blood samples).

Finally, if you have a large dataset and think you could have batch effects, there is a recent paper that applies a batch effect correction method (RUVm) in top of Funnorm: http://nar.oxfordjournals.org/content/early/2015/05/18/nar.gkv526.full

Hope this helps,

Jean-Philippe

[niutster]

Hi, I'm working on 450 HM data,too but i've got some additional data files in .txt format such as "TableControl annotations HM450" and "TableControl data no normalisation HM450 "and i dont know what are they. Do you have any idea?

Thanks

Answer 2

[yourA]

I've used wateRmelon, it offers a selection of methods that can be ranked by performance using a series of control metrics.

http://www.ncbi.nlm.nih.gov/pubmed/23631413

Answer 3

[Shicheng Guo]

More than 8 normailzation methods were proposed for HM450K array, including QN, PBC, SWAN, SQN, Dasen, BMIQ, ASMN, and QN.BMIQ. The best normalization method should be dependent your motivation, study design and data pre-processing (DMS, DMR, EGWAS). In my opinion, set some positive control and try each normalization method one by one and choose the best one which could explained and fit your postive control.