[BioC] Removing Batch effect with Limma

[Ron Ophir]

Hi,
I have the following MA experiment running on Affymetrix technology. 
It includes 4 condition and 3 time points for each condition as
follow:
FileName	Targets	Time	Treatment	Batch
F21.CEL	I18	18	I	3
F22.CEL	I18	18	I	3
F25.CEL	N18	18	N	3
F26.CEL	N18	18	N	3
F23.CEL	R18	18	R	3
F24.CEL	R18	18	R	3
F19.CEL	S18	18	S	3
F20.CEL	S18	18	S	3
F13.CEL	I24	24	I	2
F39.CEL	I24	24	I	4
F9.CEL	I24	24	I	1
F10.CEL	N24	24	N	1
F14.CEL	N24	24	N	2
F41.CEL	N24	24	N	4
F12.CEL	R24	24	R	2
F40.CEL	R24	24	R	4
F8.CEL	R24	24	R	1
F11.CEL	S24	24	S	2
F38.CEL	S24	24	S	4
F7.CEL	S24	24	S	1
F43.CEL	I28	28	I	4
F56.CEL	I28	28	I	5
F57.CEL	I28	28	I	5
F49.CEL	N28	28	N	4
F60.CEL	N28	28	N	5
F61.CEL	N28	28	N	5
F48.CEL	R28	28	R	4
F58.CEL	R28	28	R	5
F59.CEL	R28	28	R	5
F46.CEL	S28	28	S	4
F54.CEL	S28	28	S	5
F55.CEL	S28	28	S	5

The replicates in this experiment are biological replicates not
technical. Note that some of the replicates were prepared in a different
batches.
>From the attached PCA image, it seems we have here a batch effect
(lines are connection among replicates), i.e., is highly contributes to
the variance. Would it be correct to run
duplicateCorrelation(MA, design, ndups = 1, block = targets$Batch)
before lmFit()
to remove the batch effect or it is not right to do so because the
replicates (e.g., 28 R 4 and 28 R 5) are not technical?
If it is wrong, are you suggesting in the future to double the number
of chips and preparing one technical replicate in one batch and another
replicate in other batch?
Thanks,
Ron