[BioC] Design/Contrast for Two-Channel Experimental Setup

[Gordon K Smyth]

Dear Joseph,

On Mon, 6 Jan 2014, Joseph Shaw wrote:

> Dear Gordon,
>
> Thank you very much for your response!
>
> I have two brief follow-on questions pertaining to your previous mail.
>
> On Sun, Jan 5, 2014 at 10:39 PM, Gordon K Smyth <smyth at wehi.edu.au> wrote:
>>
>>> Date: Sat, 4 Jan 2014 19:58:32 +0000
>>> From: Joseph Shaw <josph.sh at gmail.com>
>>> To: Ryan <rct at thompsonclan.org>
>>> Cc: bioconductor at r-project.org
>>> Subject: Re: [BioC] Design/Contrast for Two-Channel Experimental Setup
>>>
>>> It was my belief that the experimental setup would imply that the dye 
>>> effect would be confounded with the biological effect - thanks for 
>>> clarifying that this is indeed the case. However, I'm still slightly 
>>> confused about the dye effect term; specifically, shouldn't the loess 
>>> normalisation (performed by normalizeWithinArrays() function) correct 
>>> for the dye effect? If this is the case, why is a dye effect term 
>>> required?
>>
>> The loess normalization done by normalizeWithinArrays() accounts for a 
>> global dye effect trend.  However it is possible that some of the 
>> probes on the array might show special dye effects specific to those 
>> probes which don't follow the overall dye effect trend.  It is the 
>> purpose of a dye effect term in the linear model to allow for the 
>> possibility of such probe-specific dye effects.
>
> Am I correct in suggesting that such a dye-effect term (assuming one 
> exists) will be represented by a model parameter (intercept) estimate 
> common to all observations in a given gene model? If this is the case, 
> this dye effect (the intercept estimate) will be applied to all 
> observations (across replicates) for a given gene model as opposed to 
> any single observation within the gene model.

Yes, each dye effect term will be applied to a row of expression values. 
Each row of data from a microarray corresponds to a microarray spot or 
probe, not necessarily a "gene model".

> As such, the dye-effect term is a gene-specific as opposed to 
> observation specific.

It is probe-specific, as I said before.  No one claimed it was 
"observation specific".

>>> Also, with a view to identifying differentially expressed genes, is the
>>> sample code provided in my previous mail otherwise correct? Are there any
>>> alterations that I should consider?
>>
>>
>> The line
>>
>>  MA.b=normalizeBetweenArrays(MA, method="quantile")
>>
>> is not needed, and is obviously superfluous in your code anyway.
>>
> Could you briefly elaborate on why this line is not needed? As I
> currently understand it, normalization between arrays is advantageous
> if there exists a disparity between replicate distributions (in which
> case a scaling procedure such as quantile normalization can be
> implemented); is this correct?

Well, first off, your code never did between-array normalization, because 
the between-array command produced a data object that was not used in the 
subsequent analysis.

Your understanding about between array normalization might be from 
experience with single channel arrays.  For two colour arrays, the loess 
normalization step already puts the M-values for different arrays on a 
common scale so there is (usually) nothing more to do.  Loess 
normalization is superior to quantile normalization because it uses the 
pairing of channel values from the same spot.  A subsequent quantile 
normalization step is not needed and would mess up the job done by loess 
normalization.

Also, I would not describe quantile normalization as a "scaling" 
procedure.

Best wishes
Gordon

> Kind regards,
>
> Joseph
>

______________________________________________________________________
The information in this email is confidential and intend...{{dropped:4}}