Be careful not to confuse terminology here. 'Background correction' of Illumina data occurs at the raw bead level, and is typically the default setting in the Illumina software. 'Background normalisation' occurs at the bead summary level, and makes use of the negative controls to try and calibrate the data between arrays (see https://stat.ethz.ch/pipermail/bioconductor/2006-March/012178.html for further discussion). Our experience is that background correction can be worthwhile, provided that it is done carefully, while background normalisation is unhelpful if you want to analyse data on the log scale because it often produces negatives. Also note that the BeadStudio software (we have version 2.3.47) has a pop-up message warning against background normalisation for expression data. At the moment we use quantile normalisation on the log2 scale to normalise BeadStudio summary data which hasn't been normalised already. You could also try rank.invariant without doing the background normalisation (I'm not sure if this is better done on the original or log2 scale?), i.e. T = apply(exprs(BSData), 1, mean) BSData.rankinv = assayDataElementReplace(BSData.bgnorm, "exprs", rankInvariantNormalise(exprs(BSData.bgnorm), T)) fit = lmFit(log2(exprs(BSData.rankinv)), design) If this fails, Sean's suggestion of replacing the negative values with small positive values (or even NA's) should work. I hope this helps. Best wishes, Matt On 15/2/07 19:14, "Lynn Amon" <lynnamon at="" u.washington.edu=""> wrote: > Do you really need to do background subtraction with Illumina data? Our > experience is that this step is not necessary. > > Lynn Amon > Research Scientist > University of Washington > > > Nieves Velez de Mendizabal wrote: >> We are analyzing some data of Illumina. There are three kind of >> normalization. First of them is the method of rank invariant >> normalization, recommended by Illumina, and we would like to apply it: >> >> >> BSData.bgnorm = backgroundNormalise(BSData) >> T = apply(exprs(BSData.bgnorm), 1, mean) >> BSData.rankinv = assayDataElementReplace(BSData.bgnorm, "exprs", >> rankInvariantNormalise(exprs(BSData.bgnorm), T)) >> >> >> But in BSData.rankinv I have negative values so I cannot apply the >> method lmFit in order to analyze the differential expression because of >> the log2 transformation applied. >> >> fit = lmFit(log2(exprs(BSData.rankinv)), design) >> >> Are these two methods (rank inv method and lmFit) incompatible? >> What kind of normalization should I use in order to search >> differentially expressed genes in micro arrays of Illumina? >> >> Thanks