[BioC] lumi for Illumina methylation data - understanding the colour adjustment

[Pan Du]

Hi Lavinia

You are right. For Infinium methylation chip, the methylated and
unmethylalted probes are always in the same color channel. When performing
non-linear color adjustment (like quantile or other curve fitting methods),
the ratio of methylated and unmethylalted probe intensities may change. For
example, the intensities in the high range may adjust differently from those
in the low intensity range, which will also cause beta or M-value change
slightly. 

As any adjustment or normalization may also bring additional bias to the
data, if the color imbalance between two color channels are consistent
across the entire dataset, we do not recommend to perform aggressive color
adjustment, like quantile adjustment. However, if the color imbalance are
inconsistent across samples, then the color adjustment becomes important.

Hope this clarifies your question.


Pan


On 12/20/10 10:20 PM, "Lavinia Gordon" <lavinia.gordon at mcri.edu.au> wrote:

> Hi Pan
> 
> Thank you for your reply, very prompt at this busy time of year!
> I think I am getting confused with how the HumanMethylation27 array
> works, or perhaps I don't understand the colour adjustment properly.
> My query is really, if the methylated and unmethylated are the same
> colour, and that colour is adjusted (so I assume Cy3 is adjusted
> differently to Cy5, as Cy5 is incorporated differently), why are the
> intensities of the methylated adjusted differently to the intensities of
> the unmethylated if they are the same colour?
> 
> Many thanks,
> 
> Lavinia.
> 
> On 21/12/2010 2:56 PM, Pan Du wrote:
>> Hi Lavinia
>> 
>> It is hard to answer your question without knowing the quality of your data.
>> My suggestion is performing visually color bias check first before applying
>> color bias adjustment. If the color bias is not severe, then only perform
>> conservative adjustment (scaling and shift adjust) or not perform any color
>> adjustment at all. Need to know the smooth quantile color adjustment has
>> strong assumption of the data (same distribution of two color channels).
>> Quantile normalization may bring bias after adjustment, this is the same as
>> the expression microarray normalization. If you would like to, please send
>> me the plot produced by plotColorBias2D of this sample. Also, I will add
>> more detailed description of this in the vignette.
>> Thanks for reporting this!
>> 
>> 
>> Pan
>> 
>> On 12/20/10 7:06 PM, "Lavinia Gordon"<lavinia.gordon at mcri.edu.au>  wrote:
>> 
>>> Dear Dr Du
>>> 
>>> I am a big fan of /lumi/ and was delighted to see that you have made it
>>> compatible with methylation arrays.  I have used these new functions on
>>> several of my datasets and am very happy with the alternative method of
>>> working with M values.  I just have one query regarding the colour
>>> adjustment.
>>> 
>>> So, for example, probe A is a red probe, and has a (GenomeStudio)
>>> unmethylated intensity of 2205 and a methylated intensity of 2822.
>>> 2822/(2822+2205)
>>> beta = 0.5613686
>>> After colour adjustment, it has an unmethylated intensity of 1718.882
>>> and a methylated intensity of 2576.6539:
>>> 2576.6539/(2576.6539+1718.882)
>>> beta = 0.5998446
>>> 
>>> Why, if the methylated is the same colour as the unmethylated, has the
>>> unmethylated intensity decreased by 23% but the methylated by only 9%?
>>> 
>>> with thanks for your time,
>>> 
>>> Lavinia Gordon.
>> 
>> 
>> 
>> 
>> 
>