Dear Efstathie,

In response to your questions:

1) Regarding lumiB, the default method is 'none', which assumes that the BeadStudio output data is background corrected. So by default, it will do nothing. I did not want that because I did not background correct my data in Genome Studio. I tried using the bgAdjust method but this resulted in some negative values that could not be handled by the following log2 transformation. Thus, I ended up using the 'forcePositive' method, which prepares its output for log2 transformation.

2) Regarding my filtering criteria at the linear model fit step (both after neqc and after lumi preprocessing) I used:

ct<-factor(targets$T790M) #the mutation I am interested in - I have named the columns of my matrix accordingly
design <- model.matrix(~0+ct+targets$Batch)
colnames(design) <- c(levels(ct), 'batch')
arrayw <- arrayWeights(gefR, design)
dupcor <- duplicateCorrelation(gefR, design, block=targets$Sample, weights=arrayw) #correlation between the samples
fit <- lmFit(gefR, design, block=targets$Sample, correlation=dupcor$consensus.correlation, weights=arrayw)
contrasts <- makeContrasts('pos-neg',levels=design) #I am interested in the contrast: mutation-positive versus mutation-negative
fit2 <- contrasts.fit(fit, contrasts)
fit2 <- eBayes(fit2, robust=TRUE, trend=TRUE) #I added the trend=TRUE after your suggestion
top.tab = topTable(fit2, adjust='fdr', number=Inf, p.value=0.05) #

At the last step: After neqc, if I additionally put lfc=1, I won't get any significant genes. Using just p-value criteria, I get 20 DE genes. After lumi preprocessing, on the contrary, I am getting 60 DE genes when I apply both p-value and lfc criteria.

My concern (and the source of all these differences) is the big difference in the truly expressed genes I remain with, after each p-value detection filtering. As I wrote at the original post, when using **y$other$Detection < 0.01** I get ~4,000 truly expressed probes, whereas when I use **detectionCall()** function of lumi I get ~21,000 truly expressed probes. So, much larger a dataset to work with and, obviously, more DE genes detected at the end...

Thank you very much!