EPIC arrays: best normalization method?

[Shraddha Pai]

Hi everyone,

Jean-Phillippe, I'm hoping you can provide advice regarding this matter.

I have been using functional normalization with noob, via the minfi package, to normalize our EPIC arrays. 

To my knowledge, among all normalization approaches, functional normalization has the best evidence to date for removing technical variation in the Illumina BeadChip methylome platform (http://www.ncbi.nlm.nih.gov/pubmed/25599564), and for maximizing replicability across batches. More recently, noob was recommended as the sole strategy to use if considering merging 450K with EPIC arrays (http://biorxiv.org/content/early/2016/07/23/065490).

A collaborator recently attended a talk on the TCGA methylome chips and took away the message that "noob is the best normalization method for EPIC arrays, it both normalizes the arrays and eliminates batch effect. It also better to do minimal normalizing apart from Noob, unless the batch effects are very, very apparent."

I am wondering how to assess this comment. Functional normalization appears to get covariates from the principal components of the control probes, and the paper clearly states that providing experimental design information (i.e. batch information) is not required. Is there more recent evidence suggesting that functional normalization should be limited to some types of datasets? Also I didn't think noob removed technical variation beyond background subtracting and dye bias correction.

 

Your guidance would be appreciated, because I haven't found any literature to this effect.

Thanks,

Shraddha

----
Shraddha Pai
Postdoctoral Fellow, http://baderlab.org, University of Toronto
Affiliate Scientist, The Centre for Addiction and Mental Health, Toronto

[Tim Triche, Jr.]

I'm not JPF, but...

Dr. Shen said that because she's been working on testicular germ cell tumors.  Since these tumors break nearly all assumptions about DNA methylation in humans, they also break normalization methods that assume the experimenter is dealing with human samples.  Noob doesn't really assume anything other than "a bigger sample is better than a smaller sample" and "unbiased is better than biased, with equal variance".  To the best of my knowledge, there aren't any interesting samples that have been run on all three platforms (hm27, hm450, hmEPIC) the way that the TCGA AMLs were on hm27 and hm450 (since there were 192 of them, it was pretty easy to benchmark methods using those).  There will be, though, in the next few months, and they're from samples that I personally feel are irreplaceable. So then we'll have a real benchmark.

I disagree that noob gets rid of batch effects by itself.  It may minimize them, but depending on the experiment, it may not be sufficient.  There are several useful batch effect correction methods (to include funnorm, pSVA, and ComBat) which may or may not be useful in that respect. 

Back to JPF (err, Dr. Fortin), now. 

--t

[Jean-Philippe Fortin]

Hi Shraddha, Hi Tim,

In my experience, when breaking down the effect of the two steps in noob, that is (1) background correction and 2) dye bias correction, step (2) is the step that helps a lot for between-sample comparison; the dye bias between red and green is very batch dependent, so the step 2) really does remove a lot of batch effects in that sense, but at a global level (simple scaling factor between the two color channels). Batch effects can be local in the sense that different cpgs/probes can be affected differently by the batch, so as Tim said, methods like SVA, pSVA, RUV or ComBat should be used/explored after normalization. Funnorm is also a global normalization method fixing the marginal distributions, so it fixes "global" batch effects, and will not necessarily take care of the local batch effects. Here is a summary, glossary and a guideline for normalization /  batch effects removal; Tim, let me know if you agree:

- Batch effects can be broken down into a global effect (affecting all the probes) and local effects, that is probe-specific batch effects. noob and any good normalization will get rid of the global effect, but batch effects removal tools (SVA, ComBat, RUV) should be used to remove any residual local effect. 

- In my experience, the noob steps always improves downstream analysis, so should always be performed. Functional normalization includes noob as an internal first step by default, and removes additional global effects as shown in the paper, so IMO should always be run as well, especially with large sample sizes. In our recent paper (http://biorxiv.org/content/early/2016/07/23/065490), funnorm wasnt adding much to noob only, but I am currently working on significant extensions of the algorithm that does improve its performance.

Hope this helps,

Jean-Philippe

[Tim Triche, Jr.]

This is pretty much my experience as well.  Happy to  hear that funnorm is getting some updates -- it's a great foundation to build upon, as long as experimenters recognize the assumptions which must be made in order to get better control of non-uniform technical artifacts.  Noob is a very "dumb" method that fixes as much as it can by brute force (enlarging the sample size for background correction, which tends to correct probe design differences, and then reciprocating the dye bias to remove as much as possible batch-related dye bias effects).  The trouble with "smart" methods is that sometimes there are generally-valid assumptions baked in, which can fail on corner cases.  It's important for an experimenter to be aware of this -- making assumptions is not a design flaw! It's how you get better performance in the common case.

[Shraddha Pai]

Hi Tim & Jean-Philippe,

Thank you for the detailed discussion of sources of technical variation, and in clarifying exactly which aspects noob and funnorm correct for.  (Thank you also for developing these methods!) 

I would have thought that a method that estimates non-biological covariates from control probes (my understanding of funnorm) would be applicable in any experiment, because control probes should be unaffected. But perhaps not so, as evidenced by:

1)  the distortion of methylation estimates in testicular germ cell tumours (empirically found to perform poorly, by comparing normalized estimates to known positive/negative controls, presumably?)

2) this post from Jean-Philippe on biostars, suggesting that quantile norm may be the more appropriate choice in EWAS-like scenarios (closer to my project), or situations where one is not expecting large differences between groups as in cross-tissue comparison (https://www.biostars.org/p/149628/). Which in turn begs the question of why a “blanket” quantile normalization and why not something like SWAN.

In my situation, I will likely start with (noob+funnorm) and then run through the same analysis with noob+quantile for the EWAS study, see if there are major differences in the overall results. In both scenarios, I will follow up with SVA and ComBaT for batch correction.

Thanks again,

Shraddha