Dear Yannick, Most clustering algorithms are invariant relative to linear transformations, so you can cluster with your individual array data without any transformation. It doesn't matter than you don't have a common reference, except from the point of view of interpretting the clustered patterns that you end up with. (Personally I usually prefer to cluster on fitted coefficients, but that's another matter.) To visualize variability, you have 10 reps of T0-T1, 10 of T1-T2, 10 of T2-T10. Why not just do boxplots for each group. Again no need to transform. Best wishes Gordon >Date: Thu, 9 Aug 2007 18:58:46 +0200 >From: Yannick Wurm <yannick.wurm at="" unil.ch=""> >Subject: [BioC] Loop design -> inferring >To: bioconductor at stat.math.ethz.ch > >Hello list, > >I am a graduate student doing some micorarray experiments on social >behavior in ants. I've recently started using limma, and have been >pleasantly surprised by the elegance of its implementation. This is >my first question to the list. > >My experiment is a 3 point time-course that I set up as a loop design >on our two-color cDNA arrays: > T0 -> T1 -> T2 ---> (back to T0) >My 10 replicates of this are dye-balanced. > >My favorite contrasts are T1-T0 and T2-T0. I can get my genes >estimated relative expression levels through either topTable or by fit >$coefficients. >However, I would like to visuallze relative expression levels, at the >level of individual replicates. That way I can get visual feeling of >how much variability there is between my replicates. And do >clustering as if I had used a reference design. >So for each gene I want 10 values for T1-T0, and 10 values for T2-T0. >I could get these by simply taking the numbers from my direct >comparisons. But it feels wrong ignoring the information provided >indirectly about T1-T0 through (T1-T2)+(T2-T0). > >How would you go about this? > >Thanks for any tips, > >Yannick

Hello Gordon & list, thanks for your help. Clustering on the fitted coefficients make sense to see a gene's general profile. But once again, getting a feeling for a gene's variability is important because it gives information about both the gene's robustness, as well as the sample's. And if I'm looking at a large number of genes, it doesn't make sense to do boxplots for each. What I did for now is just directly use my individual array data. But I think I'd get a better estimation if I used the information provided by the loop. So for each comparison, I think I'll use a weigted average of the 2 loop directions. Eg if I want to estimate differences between T0 and T1, I could do: ratio = 2/3 * (T1vsT0) + 1/3 * (T1vsT2 - T2vsT0) Kind regards, yannick On Aug 12, 2007, at 11:41 PM, Gordon Smyth wrote: > Dear Yannick, > > Most clustering algorithms are invariant relative to linear > transformations, so you can cluster with your individual array data > without any transformation. It doesn't matter than you don't have a > common reference, except from the point of view of interpretting the > clustered patterns that you end up with. > > (Personally I usually prefer to cluster on fitted coefficients, but > that's another matter.) > > To visualize variability, you have 10 reps of T0-T1, 10 of T1-T2, 10 > of T2-T10. Why not just do boxplots for each group. Again no need > to transform. > > Best wishes > Gordon > >> Date: Thu, 9 Aug 2007 18:58:46 +0200 >> From: Yannick Wurm <yannick.wurm at="" unil.ch=""> >> Subject: [BioC] Loop design -> inferring >> To: bioconductor at stat.math.ethz.ch >> >> Hello list, >> >> I am a graduate student doing some micorarray experiments on social >> behavior in ants. I've recently started using limma, and have been >> pleasantly surprised by the elegance of its implementation. This is >> my first question to the list. >> >> My experiment is a 3 point time-course that I set up as a loop design >> on our two-color cDNA arrays: >> T0 -> T1 -> T2 ---> (back to T0) >> My 10 replicates of this are dye-balanced. >> >> My favorite contrasts are T1-T0 and T2-T0. I can get my genes >> estimated relative expression levels through either topTable or by >> fit >> $coefficients. >> However, I would like to visuallze relative expression levels, at the >> level of individual replicates. That way I can get visual feeling of >> how much variability there is between my replicates. And do >> clustering as if I had used a reference design. >> So for each gene I want 10 values for T1-T0, and 10 values for T2-T0. >> I could get these by simply taking the numbers from my direct >> comparisons. But it feels wrong ignoring the information provided >> indirectly about T1-T0 through (T1-T2)+(T2-T0). >> >> How would you go about this? >> >> Thanks for any tips, >> >> Yannick > > _______________________________________________ > Bioconductor mailing list > Bioconductor at stat.math.ethz.ch > https://stat.ethz.ch/mailman/listinfo/bioconductor > Search the archives: http://news.gmane.org/ > gmane.science.biology.informatics.conductor