[BioC] limma question

[Gordon Smyth]

At 04:37 PM 20/06/2003, leon ding wrote:
>I am currently evaluating difference between one-color and  two-color 
>hybridization to decide which one should be used for upcoming project. For 
>2-color, I found in my self vs self hybs, the gene-specific cy5/cy3 bias 
>(i.e., reproducible red or green spots on self-self slides) is quite 
>significant. This will affect the real sample-control experiments 
>especially if the fold changes of differential _expression is small. A 
>compensate solution is to divide the ratio of sample-control slides to 
>control-control slides. Although there is no reference RNA samples, the 
>"Two-Sample Experiments" medel in limma may fit very well for this 
>purpose. ( I appreciate to hear your opiniones on this).An alternative is 
>to use single color hyb, one is control and another is sample, and this 
>will cut half of the labor and cost. But can limma handle one-color 
>experiments? If not, is there any good packages for single color 
>normalization and significance analysis? My guess is that it will have to 
>involve scaling and wonder if it can generate similar results as 2-color 
>experiment.

The LIMMA functions will handle one-color experiments just as easily as 
two-color. One-color experiments are actually the same as Affymetrix data 
as far as the linear model functions are concerned. Look at Section 10 of 
the LIMMA User's Guide and then go back to Section 6. Here is an example:

Suppose that 'E' is a matrix containing your background-corrected 
intensities with rows corresponding to genes and columns to arrays. Suppose 
that the first two arrays are controls and the next three are sample 
arrays. I would analyse them like this:

E <- pmax(E,1)    # need to avoid negative intensities if you are to use 
quantile normalization
E <- log(E,2)       # convert to log-2 scale
E <- normalizeQuantiles(E)       # quantile normalization of single-channel 
intensities
design <- cbind(Control=c(1,1,1,1,1),SamplevsControl=c(0,0,1,1,1))
fit <- lm.series(E,design)
eb <- ebayes(fit)

The second column of eb$t contains moderated t-statistics for differential 
expression between the sample and control samples.

toptable(coef=2,fit=fit,eb=eb)

will display the top genes most likely to be differentially expressed.

I am very surprised that you find single channel analysis more accurate 
than two-color, but that is another matter.

Gordon

>Leon