[BioC] Another question about normalization of data

[Kasper Daniel Hansen]

On Wed, Jul 4, 2012 at 4:54 AM, Gustavo Fernández Bayón
<gbayon at gmail.com> wrote:
> Hi everybody.
>
> Not so long ago, I asked in this list about some normalization issues. The question and its very interesting replies, from which I have learned a lot, can be found here:
>
> http://comments.gmane.org/gmane.science.biology.informatics.conductor/41812
>
> It seems to me that, the more I am getting into Bioinformatics, the less I know about everything. I usually doubt about everything, and I am always asking, step by step, if I am doing things correctly.
>
> Now, I want to test some ideas on a 450K methylation array data base. Main idea is to try to classify probesets in families according to their behavior with respect to some phenotype variables. I have several ideas I would like to try on this data, and the first step has been to import, review, visualize and try to understand the global structure of the beta values I have at my hand.
>
> Once loaded, I have made two box plots. One shows the distribution of beta values among the 40 samples, and the other shows the distribution among the first 100 probesets.
>
> I have shared the plots at my Google Docs account:
>
> https://docs.google.com/open?id=0Bw-_OWjrT9U4cTlZblR0UkVhWG8
> https://docs.google.com/open?id=0Bw-_OWjrT9U4d3FTZTQtNWJFUVE
>
>
>
> My question might sound stupid, but I want to deeply understand what is going on with these plots.
>
> For the beta vs. probeset:

probesets is a pretty bad word to use here, I think.

> - I guess the variability is normal. Some probes are methylated most of the time, some not, and there are a lot of differences in their behavior. This is the common behavior, isn't it?

Yes, of course

> - Boxplot might not be the best solution here, because the distribution need not to be unimodal, I think. Am I right?

Probably

> - My intuition is that these values should be normalized in case we were going to use something like SVM-RFE to do probeset selection. Again, is my intuition right?

Well, there is no way around the fact that different CpGs will have
vastly different behavior.  If your model selection procedure needs
similar distributions for each feature, you will need to use a
different selection tool.


> For the beta vs. sample:
>
> - Data distribution seems more regular than in the other plot. Is that an effect of the underlying normalization that GenomeStudio does? Or is the way beta values across samples are supposed to behave?

This is the consequence of now looking at the marginal behavior across
the genome (or at least the part of the genome assayed by 450k).

> - Although they seem regular, there are still small differences among medians, which made me think. Would a normalization of this data benefit following experiments?

Perhaps.  Note that the - depending on the question and the samples -
you cannot assume that these marginal distributions are the same.  For
example, in cancer they are know to be very different. See for example
  Hansen, K. D. et al. Increased methylation variation in epigenetic
domains across cancer types. Nat Genet 43, 768–775 (2011).
This means you probably have to be smart, like
  Aryee, M. J. et al. Accurate genome-scale percentage DNA methylation
estimates from microarray data. Biostatistics 12, 197–210 (2011).
Unfortunately, it is not clear that this trick (using many background
probes) can be done on 450k.  But perhaps you can assume that these
marginal distributions are similar in your experiment.

> In general, I would like to know if the plots show a normal behavior, if it is the expected one, or if I should kind of normalize them using a predefined or standard method.

This is still an open problem.  You should do whatever normalization
helps you to improve signal to noise in the context of your analysis.
Of course, this is a general observation that doesn't help you much

Kasper