[BioC] Dye effect problem with Limma package

[Gordon K Smyth]

Dear Jean-Laurent,

You say that "theoretically" the coefficients for KO and WT should be 
unchanged when a dye effect is added to the model.  This is not true if 
your MA object contains weights or missing values.

For the design matrix you give, limma does indeed give unchanged 
coefficients in the absence of weights, missing values, blocking, and so 
on.

Best wishes
Gordon

> Date: Mon, 16 Mar 2009 16:35:16 +0100
> From: Jean-Laurent Ichant? 	<jean-laurent.ichante at cgm.cnrs-gif.fr>
> Subject: [BioC] Dye effect problem with Limma package
> To: bioconductor at stat.math.ethz.ch
>
> Dear All,
> I have one question that makes me perplex, and no message concerning
> this problem is described in the mailing list.
> I'm using the Limma package for the comparison of two experimental
> conditions (KO vs WT) in a 'reference design' with dye-swap.
> My experimental design is the following:
>
>                        Cy5      Cy3
> Array1              Ref       WT
> Array2              Ref       WT
> Array3              Ref       WT
> Array4              WT       Ref
> Array5              WT       Ref
> Array6              WT       Ref
> Array7              Ref       KO
> Array8              Ref       KO
> Array9              Ref       KO
> Array10            KO       Ref
> Array11            KO       Ref
> Array12            KO       Ref
>
> In such design it is possible to estimate a dye effect.
>
> First, I have fitted a very simple model.
>
>>  fit1 <- lmFit(MA, design=design1)
> where design1 is the following design matrix:
>
>>  design1
>                                   KO       WT
> 208_REF-WT.1B           0          -1
> 213_REF-WT.25B         0          -1
> 301_REF-WT.3B           0          -1
> 210_WT.2B-REF           0          1
> 215_WT.26B-REF         0          1
> 304_WT.5B-REF           0          1
> 209_REF-KO.28B         -1         0
> 214_REF-KO.27B         -1         0
> 303_REF-KO.13B         -1         0
> 211_KO.18B-REF         1          0
> 300_KO.15B-REF         1          0
> 302_KO.17B-REF         1          0
>
>
> Then I have fitted a second model including dye effect.
>
>>  fit2 <- lmFit(MA, design=design2)
> where design2 is the following design matrix:
>
>>  design2
>                                   DyeEffect          KO       WT
> 208_REF-WT.1B          1                      0          -1
> 213_REF-WT.25B         1                      0          -1
> 301_REF-WT.3B          1                      0          -1
> 210_WT.2B-REF          1                      0          1
> 215_WT.26B-REF         1                      0          1
> 304_WT.5B-REF          1                      0          1
> 209_REF-KO.28B         1                      -1         0
> 214_REF-KO.27B         1                      -1         0
> 303_REF-KO.13B         1                      -1         0
> 211_KO.18B-REF         1                      1          0
> 300_KO.15B-REF         1                      1          0
> 302_KO.17B-REF         1                      1          0
>
>
> In theory fitted values should be identical for the two coefficients (KO
> and WT), but in my case observed values are different (for certain genes).
>
>>  fit1$coef[1:5,]
>             KO          WT
> [1,] -0.8372489 -0.19457588
> [2,] -0.3208879 -0.15328678
> [3,] -0.7202544 -0.69133039
> [4,] -0.1262356 -0.08165633
> [5,] -0.2568482 -0.18354345
>
>>  fit2$coef[1:5,]
>        DyeEffect         KO          WT
> [1,] -0.411091881 -0.8372489 -0.27679426
> [2,] -0.007895425 -0.3182560 -0.15328678
> [3,] -0.004396641 -0.7202544 -0.69133039
> [4,]  0.200046876 -0.1262356 -0.08165633
> [5,] -0.052855276 -0.2568482 -0.19411451
>
>
> What is the origin of these differences?
> Does anybody have any suggestions?
>
> Thanks again
> Jean-Laurent