Hello all, like the topic already says, I am trying to use spike-ins for normalization purposes. I ran two Agilent miRNA microarrays and when I plot the raw data I see slight variation between the two arrays, meaning the values of Array 2 are around 1.3-fold higher than the ones from Array 1. Therfore I decided to do normalization. I already read a lot about loading the data into BioC and some preprocessing strategies. Since there is still a large debate about the right preprocessing strategy I decided to figure out which of them is working best for me. I used limma so far for loading the data and for some preprocessing, but I also loaded the data by read.table() and then used vsn() and other. Unfortunately after normalization using the implemented strategies (quantile, vsn) all differentially expressed genes are gone, meaning that values which have been different by 10-fold and more in the raw dataset, are now almost equal. Therfore I decided to do spike.in normalization. The array has several "control" spots as well as negative and positive controls. I read a lot about normalization with spike-ins but I just can't figure out how to do it. I know that vsn2() offers this by fitting a transformation only to a small number of probes and than normalizing the bulk data "towards" this transformed data. However, if I extract the supposed spike-ins from my data and store it in a new matrix, doing the transformation just on these and storing the transformed values in a vsn object, I can't do the transformation of the remaining genes on this vsn object. > fit=vsn2(mrawObj_TGS.spikeins, lts.quantile = 0.7) vsn: 912 x 32 matrix (1 stratum). 0% done. 100% done. Please use 'meanSdPlot' to verify the fit. > fit2=vsn2(mrawObj_TGS.genesubset, fit, lts.quantile = 0.7) Fehler in vsnMatrix(y, reference, strata, ...) : 'nrow(reference)' must be equal to 'nrow(x)'. I understand that the matrix containing the transformed values (reference) mus have the same number of rows as the object I wan't to be transformed. But there is just no way for me to get to that point since I only have around 200 spikes but more than 12.000 genes. I also fitted a vsn transformation on the data from Array 1 and used these transformation for vsn transformation of Array 2 to get rid of the between array normalization. But I don't now if this is the right thing to do! So my question is if anybody already has experience in normalization using spike-ins and how your advice would be what I shall do now. Thanks a lot in advance! Christian Eisen

Dear Christian Two points: 1.) Please have a look at the "predict" method for "vsn" objects, this does what you want. library("vsn") method ? predict("vsn") E.g.: data("kidney") spikeInOnly = sample(nrow(kidney), 100) fit = vsn2(kidney[spikeInOnly, ]) nkid = predict(fit, kidney) 2.) Perhaps you need to face the possibility that the arrays just do not detect any differential expression (e.g. that it is swamped out by background signal?) - let us know! Best wishes Wolfgang ------------------------------------------------------------------ Wolfgang Huber EBI/EMBL Cambridge UK http://www.ebi.ac.uk/huber 21/10/2008 07:44 Christian Eisen scripsit > Hello all, > > like the topic already says, I am trying to use spike-ins for > normalization purposes. > I ran two Agilent miRNA microarrays and when I plot the raw data I see > slight variation > between the two arrays, meaning the values of Array 2 are around > 1.3-fold higher than > the ones from Array 1. Therfore I decided to do normalization. > I already read a lot about loading the data into BioC and some > preprocessing strategies. > Since there is still a large debate about the right preprocessing > strategy I decided > to figure out which of them is working best for me. > I used limma so far for loading the data and for some preprocessing, but > I also > loaded the data by read.table() and then used vsn() and other. > Unfortunately after normalization using the implemented strategies > (quantile, vsn) > all differentially expressed genes are gone, meaning that values which > have been > different by 10-fold and more in the raw dataset, are now almost equal. > Therfore I decided to do spike.in normalization. > > The array has several "control" spots as well as negative and positive > controls. > I read a lot about normalization with spike-ins but I just can't figure > out how to do it. > I know that vsn2() offers this by fitting a transformation only to a > small number of > probes and than normalizing the bulk data "towards" this transformed data. > However, if I extract the supposed spike-ins from my data and store it > in a new > matrix, doing the transformation just on these and storing the > transformed values in > a vsn object, I can't do the transformation of the remaining genes on > this vsn object. > >> fit=vsn2(mrawObj_TGS.spikeins, lts.quantile = 0.7) > vsn: 912 x 32 matrix (1 stratum). 0% done. > 100% done. > Please use 'meanSdPlot' to verify the fit. >> fit2=vsn2(mrawObj_TGS.genesubset, fit, lts.quantile = 0.7) > Fehler in vsnMatrix(y, reference, strata, ...) : > 'nrow(reference)' must be equal to 'nrow(x)'. > > I understand that the matrix containing the transformed values > (reference) mus have the same > number of rows as the object I wan't to be transformed. > But there is just no way for me to get to that point since I only have > around 200 spikes but more than 12.000 genes. > I also fitted a vsn transformation on the data from Array 1 and used > these transformation > for vsn transformation of Array 2 to get rid of the between array > normalization. > But I don't now if this is the right thing to do! > > So my question is if anybody already has experience in normalization > using spike-ins and > how your advice would be what I shall do now. > > Thanks a lot in advance! > > Christian Eisen >