In replicates of 3, I have a set of control and 4 treatment arrays. My question is, do I normalize (via rma) using all of the arrays at once, or do I normalize by treatment. I have asked other reliable sources and have received conflicting responses. I feel the issue is that, in one case, I may, if not completely wipe out, severely diminish any possible treatment effects, whereas, in the other case, I may actually induce a treatment effect. Any thoughts? Thank you so much, Leanna

you;ll probably get conflicting responses here too... in my experience the risk of inducing effects is much greater than the risk of diminishing effects. look at unnormalized data from replicate arrays and you will see large differences. this means that when you see large differences you cant be sure if its artifcact/obscuring variation or real biological variation. if in your experiment you expect global gene expression to be distributed roughly the same across conditions then quantiles normalization (the default on rma) will be fine. if you expect most gene expression no to change across condition then most normalizations available in the affy package (qsplines, loess, contrasts, etc...) should work fine as well. a paper by bolstad et al. in bioinformatics (2003) has some empirical results on all this. On Mon, 16 Jun 2003, Leanna House wrote: > In replicates of 3, I have a set of control and 4 treatment arrays. My > question is, do I normalize (via rma) using all of the arrays at once, or > do I normalize by treatment. I have asked other reliable sources and have > received conflicting responses. I feel the issue is that, in one case, I > may, if not completely wipe out, severely diminish any possible treatment > effects, whereas, in the other case, I may actually induce a treatment > effect. Any thoughts? > > Thank you so much, > Leanna > > _______________________________________________ > Bioconductor mailing list > Bioconductor@stat.math.ethz.ch > https://www.stat.math.ethz.ch/mailman/listinfo/bioconductor >

These are my two cents to the question. I would at first normalize everything together via rma, which does poses risks of diminishing or enhancing treatment effects. But I believe this would allow you to see the overall picture of the data and give you a general sense of how each treatment compares to each other and the control. And if you wanted to spend some more time, you could then try creating various data sets by using rma to normalize data between each treatment and the control and have 4 groups of data (each treatment normalized w/ the control). This would be a more specific analysis leaving out the enhancing or diminishing effects of normalization between many different treatments. Richard Park -----Original Message----- From: Leanna House [mailto:house@stat.duke.edu] Sent: Monday, June 16, 2003 10:25 AM To: bioconductor@stat.math.ethz.ch Subject: [BioC] normalize by or across all treatments In replicates of 3, I have a set of control and 4 treatment arrays. My question is, do I normalize (via rma) using all of the arrays at once, or do I normalize by treatment. I have asked other reliable sources and have received conflicting responses. I feel the issue is that, in one case, I may, if not completely wipe out, severely diminish any possible treatment effects, whereas, in the other case, I may actually induce a treatment effect. Any thoughts? Thank you so much, Leanna _______________________________________________ Bioconductor mailing list Bioconductor@stat.math.ethz.ch https://www.stat.math.ethz.ch/mailman/listinfo/bioconductor

Leanna House wrote: >In replicates of 3, I have a set of control and 4 treatment arrays. My >question is, do I normalize (via rma) using all of the arrays at once, or >do I normalize by treatment. I have asked other reliable sources and have >received conflicting responses. I feel the issue is that, in one case, I >may, if not completely wipe out, severely diminish any possible treatment >effects, whereas, in the other case, I may actually induce a treatment >effect. Any thoughts? > > I normalize across all experiments (and using RMA) despite the fact as you mentioned you may reduce any treatment effect. My only suggestion (which could be already too late) is to design the experiment in such a way that you can utilize two-way ANOVA with random block effect. What is your experimental design? I >Thank you so much, >Leanna > >_______________________________________________ >Bioconductor mailing list >Bioconductor@stat.math.ethz.ch >https://www.stat.math.ethz.ch/mailman/listinfo/bioconductor > >