[BioC] two pair dye-swap (replicates) conducted in different labs

[Gordon Smyth]

At 06:37 AM 27/09/2005, kevin Lin wrote:
>Dear Gordon and other users,
>
>Based on Limma usersguide and the hint you sent me, "the  #of columns in 
>designmatrix is equal to the #of RNA sources-1 for two-color direct 
>comparisons". This is what I did using command modelMatrix.

This advice, and the use of modelMatrix() itself, actually applies only to 
connected designs. Your design is unconnected. You need to assemble the 
design matrix by hand.

The solution is quite simple, and I've answered virtually the same question 
on this list before. Basically you just put the design matrices for the two 
connected parts of your experiment together. Rather than giving the 
solution again here, I think a longer term solution is to say that you 
really should consult one of the senior statisticians in your department. 
The MDACC employs many experienced statisticians, some of whom are experts 
in microarray analysis.

>## something i don't understand ##
>1) why my corfit$consensus is NaN? Am I supposed to get a number which 
>indicates a relation between two block arrays(first 2 and last 2).

You have no business using blocking here. The two pairs of arrays are 
reflected in your design matrix, not in the error structure.

Gordon

>2) If ignoring NaN, the result of lmFit has a message "not estimable: KO 
>UV" which is the coefficient I am most interested in ("KO UV/WT control"). 
>What is happening here?
>
>Again, my TargetsGenePix is
>>TargetsGenePix
>                                      SlideNumber 
> FileName        Cy3                  Cy5
>FG WT rep1 Sep2005      3215L FG WT rep1 Sep2005.gpr       WT control WT UV
>FG WT rep2 Sep2005      3202OL FG WT rep2 Sep2005.gpr     WT UV WT control
>FG KO rep2 Sep2005      3215OL FG KO rep2 Sep2005.gpr      KO control KO UV
>FG KO rep1 Sep2005      3202L FG KO rep1 Sep2005.gpr        KO UV KO control
>
>design <- modelMatrix(TargetsGenePix, ref="WT control")
>
>>  design
>                             KO control    KO UV    WT UV
>FG WT rep1 Sep2005          0     0     1
>FG WT rep2 Sep2005          0     0    -1
>FG KO rep2 Sep2005         -1     1     0
>FG KO rep1 Sep2005          1    -1     0
>
>corfit <- duplicateCorrelation(MAgpr, design, ndups=1, block=c(1,1,2,2))
>>  corfit$consensus
>[1] NaN
>
>>  fit <- lmFit(MAgpr, design)
>Coefficients not estimable: KO UV
>
>Thanks for your time
>
>Kevin