[BioC] Design question in LIMMA

[Nataliya Yeremenko]

Hello
This is a long letter about my efforts of analysis of data in Limma.

I have a question about the proper design of my experiment
I have 3 groups to compare: A, O, and Y.
With 5 A, 8 O, and 7 Y biological samples.
I've performed in total 28 two-color microarrays (Agilent 44K)
with a mixed number of dye-swaps.

My targets file is:
HybID    fileName    sampleID    Cy3    Cy5
1    124879.txt    YvsO    Y1    O1
2    124880.txt    OvsY    O1    Y1
3    124919.txt    YvsO    Y2    O2
4    124972.txt    OvsY    O2    Y2
5    124984.txt    YvsO    Y3    O3
6    124957.txt    OvsY    O3    Y3
7    130365.txt    YvsO    Y4    O4
8    130366.txt    OvsY    O4    Y4
9    130372.txt    YvsO    Y5    O5
10    130374.txt    OvsY    O5    Y5
11    124881.txt    AvsO    A1    O1
12    124882.txt    OvsA    O1    A1
13    124982.txt    AvsO    A2    O2
14    124983.txt    OvsA    O2    A2
15    130351.txt    AvsO    A2    O2
16    124985.txt    AvsO    A3    O3
17    124958.txt    OvsA    O3    A3
18    130352.txt    AvsO    A3    O3
21    130355.txt    AvsO    A4    O4
22    130361.txt    OvsA    O4    A4
23    130362.txt    AvsO    A5    O5
24    130363.txt    OvsA    O5    A5
19    130353.txt    AvsO    A6    O6
20    130354.txt    OvsA    O6    A6
25    130375.txt    AvsO    A7    O7
26    130376.txt    OvsA    O8    A7
27    130377.txt    AvsO    A7    O6
28    130396.txt    OvsA    O6    A7

After import of the data, normalization within and between arrays and
evaluation of diagnostic plots, the question about fitting linear model 
arised.

I didn't succeed to create proper direct design for all 3 groups.
However for separate Y vs O, and A vs O it works Ok
with the design of type:
design <- cbind(Y1vsO1 = c(-1,1,0,0,0,0,0,0,0,0),
                Y2vsO2 = c(0,0,-1,1,0,0,0,0,0,0),
                Y3vsO3 = c(0,0,0,0,-1,1,0,0,0,0),
                Y4vsO4 = c(0,0,0,0,0,0,-1,1,0,0),
                Y5vsO5 = c(0,0,0,0,0,0,0,0,-1,1))
But here I think I loose info about O6, O7 and O8 which are extra 
biological replicates.
The same is valid for A vs O and I had to exclude last three hybs.

What is your advise in that case?

I have also tried to split all data into separate channels,
so producing 56 single-channel data sets.
(The reason for that was that I have even and odd number of replicates 
for my groups mixed in hybridizations)
 >targets2 <- targetsA2C(targets)
 >u <- unique(targets2$Target)
 >f <- factor(targets2$Target, levels=u)
 >design <- model.matrix(~0+f)
 >colnames(design) <- u

It works not bad until
 >corfit <- intraspotCorrelation(MA, design)
It took a lot of time and generated 43 warnings: "exceed amount of 
iterations ...."
fit <- lmscFit(MA, design, correlation=corfit$consensus)

Than a BIG question appeared: "What is the contrasts matrix is in my case?"
cont.matrix <- 
makeContrasts("(A1+A2+A3+...)/7-(O1+O2+O3+...)/8",levels=design)
fit2 <- contrasts.fit(fit, cont.matrix)
fit2 <- eBayes(fit2)
topTable(fit2, adjust="BH", number=30, resort.by"M")

Is it correct for A vs O comparison?
I've got the table finally...
And needles to say top 10 is different from my direct design A vs O (see 
above)

Regards
Nataliya





-- 
Dr. Nataliya Yeremenko 

Universiteit van Amsterdam
Faculty of Science
IBED/AMB (Aquatische Microbiologie)
Nieuwe Achtergracht 127
NL-1018WS Amsterdam
the Netherlands

tel. + 31 20 5257089
fax  + 31 20 5257064