[BioC] Limma question

[Gordon K Smyth]

Dear Niccolo,

I have to tell you that what you claim to have observed is not possible. 
If the normalized intensities were all equal, then limma would produce 
t-stat=0 and p-value=0 for any contrast between conditions.  So it would 
seem that you've made a mistake somewhere in collating results.

Your email does not contain complete code, so there isn't any way for me 
to help you find the error.

Best wishes
Gordon

> Date: Fri, 16 Dec 2011 16:58:55 +0100
> From: Niccol? Bassani <biostatistica at gmail.com>
> To: <bioconductor at stat.math.ethz.ch>
> Subject: [BioC] Limma question
>
> Dear users,
> I'm having some troubles in figuring out what's going on in limma.
> I've got some expression data from Agilent microRNA platform, I've
> pre-processed them, and wanted to do some easy differential expression
> analysis. Out of 1368 miRNAs (no filtering performed) there are 758 of
> them which show EXACTLY the same value on all of the 24 arrays
> involved. Arrays are divided in 3 groups, 8 arrays in each group.
> Data look like this (in matrix form, first rows and columns):
>
>        LN9      LN10      LN11      LN12      LN13      LN14
> 1 12.431022 12.186179 13.136163 12.121403 12.643895 12.756163
> 2  1.137504  1.137504  1.137504  1.137504  1.137504  1.137504
> 3  1.137504  1.137504  1.137504  1.137504  1.137504  1.137504
> 4  1.137504  1.137504  1.137504  1.137504  1.137504  1.137504
> 5  1.137504  1.137504  1.137504  1.137504  1.137504  1.137504
> 6  1.137504  1.137504  1.137504  1.137504  1.137504  1.137504
>
> I specify the design matrix, and run easy differential expression code:
>
> contrasts = cbind(AvsB = c(-1,1,0),AvsC = c(1,0,-1),AvsB_C =
> c(1,-1/2,-1/2),A_BvsC = c(1/2,1/2,-1))
> contrasts
>     AvsB AvsC AvsB_C A_BvsC
> [1,]   -1    1    1.0    0.5
> [2,]    1    0   -0.5    0.5
> [3,]    0   -1   -0.5   -1.0
>
> fit = lmFit(agilent,design)
> fit.contrasts = contrasts.fit(fit,contrasts)
> test = eBayes(fit.contrasts)
>
> The strange (or absurd) thing is that invariant microRNAs appear to be
> differentially expressed throughout all of the contrasts but the last
> one!
>
> test
> $p.value
>           AvsB       AvsC     AvsB_C    A_BvsC
> [1,] 0.53958575 0.42970445 0.41866547 0.5748925
> [2,] 0.03471306 0.03471306 0.01644463 1.0000000
> [3,] 0.03471306 0.03471306 0.01644463 1.0000000
> [4,] 0.03471306 0.03471306 0.01644463 1.0000000
> [5,] 1.00000000 0.23359101 0.48667557 0.1713666
> 1363 more rows ...
>
> I've drilled into the various limma functions code, but it seems that
> there's some problem with my data, maybe some kind of
> approximation...my point is that the last contrast correctly
> identifies no microRNA differentially expressed, whereas the remaining
> 3 return me t statistic which are non 0 for invariant miRNAs!!
>
> $t
>             AvsB       AvsC     AvsB_C     A_BvsC
> [1,] 6.236028e-01 -0.8051982 -0.8249186 -0.5697255
> [2,] 2.257614e+00 -2.2576137 -2.6068677  0.0000000
> [3,] 2.257614e+00 -2.2576137 -2.6068677  0.0000000
> [4,] 2.257614e+00 -2.2576137 -2.6068677  0.0000000
> [5,] 1.588357e-14 -1.2263878 -0.7080553 -1.4161107
> 1363 more rows ...
>
> Any suggestions? I've tried to round the dataset to 4 digits but the
> problem's still there, only changes the contrast with consistently
> non-differentially expressed genes...
>
> Thanx, and merry xmas everybody (know it's early, but who knows what
> will be next...)
> Niccol?

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