Gordon, thanks for the reply. I am a little confused. How can I fit contrasts if the design (not mine btw) is unconnected? As far as I can see, the hybridizations are anlagous to those shown on page 48 of the limma guide, the section that deals with single channel analysis. Note the column names of my MA object [1] "../ampcon/mev/C0-_1stround_vs_C60-_1stround_13263536.mev" [2] "../ampcon/mev/C0-_2ndround_vs_C60-_2ndround_13263534.mev" [3] "../ampcon/mev/C60-_1stround_vs_C0-_1stround_13263533.mev" [4] "../ampcon/mev/C60-_2ndround_vs_C0-_2ndround_13263531.mev" [5] "../licl/mev/C0-_vs_C60-_13260944.mev" [6] "../licl/mev/C60-_vs_C0-_13260945.mev" There are no biological replicates. The experiment was performed by taking a pool of total RNA from C0- and C60- samples. The amount of RNA was estimated for each, and dilutions were made of the C0- and C60- total RNA. These dilutions, have artificially low total RNA (as you might get from some tissue biopsy). These samples were subjected to either 1 round (cohybe C0- and C60- on slides 1,3) or 2 rounds (cohybe C0- and C60- on slides 2,4) of amplification. The total RNA samples were hybridized on slides 5 and 6. Schematically: C0-.round1 <-----> C60-.round1 C0-.round2<------->C60-.round2 C0-.totalRNA<----->C60-.totalRNA If I create this design matrix: round_1 round_2 non_amp 1 1 0 0 2 0 1 0 3 -1 0 0 4 0 -1 0 5 0 0 1 6 0 0 -1 Then I have the log ratios of sample C60 and C0 for each protocol. But I want to compare round_1 v non_amp and round_2 v non_amp to see if they are significantly different. I am interested to see which genes show significantly different log ratios. In this case, you suggest fitting the contrasts. So, I did the following to estimate the difference in log ratios. > cont.one.non<-makeContrasts("round_1-non_amp",levels=hyp.design) > cont.one.non round_1-non_amp round_1 1 round_2 0 non_amp -1 Which gives: > as.matrix(hyp.design) %*% as.matrix(cont.one.non) round_1-non_amp 1 1 2 0 3 -1 4 0 5 -1 6 1 How does this help me? Thanks Gordon K Smyth wrote: >On Wed, July 26, 2006 8:47 pm, Andrew Mcdonagh wrote: > > >>Hi Gordon, >> >>Thanks for getting back to me, although I don't think you answered my >>question. On the surface it looks like another couple of posts relating >>to this, but that's not actually the case >> >>I think I might not have explained myself clearly. I have read the limma >>guide, and I can see how you calculate the t-statistics on the >>coefficients. The example that you give is relevant to me if the >>coefficients themselves are comparisons between protocols. Clearly from >>my design matrix the effect is not a comparison of protocols. So the >>three coefficients that I estimate are the effect of protocol 1, >>protocol 2 and protocol 3. >> >> > >Use contrasts.fit to make the coefficients the comparisons you want. > > > >>Also, the question about single channel analysis relates to this. Do you >>have any suggestions >> >>Thanks >> >>Andy >> >> > >Please send questions to the Bioconductor list. > >Best wishes >Gordon > > > > >

> Date: Wed, 26 Jul 2006 15:51:06 +0100 > From: Andrew Mcdonagh <a.mcdonagh at="" imperial.ac.uk=""> > Subject: Re: [BioC] limma advice required. Investigating the > amplification of small sample RNA > To: bioconductor at stat.math.ethz.ch > > Gordon, thanks for the reply. > > I am a little confused. How can I fit contrasts if the design (not mine > btw) is unconnected? As far as I can see, the hybridizations are > anlagous to those shown on page 48 of the limma guide, the section that > deals with single channel analysis. Note the column names of my MA object > > [1] "../ampcon/mev/C0-_1stround_vs_C60-_1stround_13263536.mev" > [2] "../ampcon/mev/C0-_2ndround_vs_C60-_2ndround_13263534.mev" > [3] "../ampcon/mev/C60-_1stround_vs_C0-_1stround_13263533.mev" > [4] "../ampcon/mev/C60-_2ndround_vs_C0-_2ndround_13263531.mev" > [5] "../licl/mev/C0-_vs_C60-_13260944.mev" > [6] "../licl/mev/C60-_vs_C0-_13260945.mev" > > There are no biological replicates. The experiment was performed by > taking a pool of total RNA from C0- and C60- samples. The amount of RNA > was estimated for each, and dilutions were made of the C0- and C60- > total RNA. These dilutions, have artificially low total RNA (as you > might get from some tissue biopsy). These samples were subjected to > either 1 round (cohybe C0- and C60- on slides 1,3) or 2 rounds (cohybe > C0- and C60- on slides 2,4) of amplification. The total RNA samples were > hybridized on slides 5 and 6. > > Schematically: > > C0-.round1 <-----> C60-.round1 > > C0-.round2<------->C60-.round2 > > C0-.totalRNA<----->C60-.totalRNA > > > If I create this design matrix: > > round_1 round_2 non_amp > 1 1 0 0 > 2 0 1 0 > 3 -1 0 0 > 4 0 -1 0 > 5 0 0 1 > 6 0 0 -1 > > Then I have the log ratios of sample C60 and C0 for each protocol. But I > want to compare round_1 v non_amp and round_2 v non_amp to see if they > are significantly different. I am interested to see which genes show > significantly different log ratios. In this case, you suggest fitting > the contrasts. So, I did the following to estimate the difference in log > ratios. > > > cont.one.non<-makeContrasts("round_1-non_amp",levels=hyp.design) > > cont.one.non > round_1-non_amp > round_1 1 > round_2 0 > non_amp -1 This does exactly what you want if you use it in conjunction with contrasts.fit(), i.e., it finds genes for which the log-ratio is different after 1 round amplification as compared to no amplication. > Which gives: > > > as.matrix(hyp.design) %*% as.matrix(cont.one.non) > round_1-non_amp > 1 1 > 2 0 > 3 -1 > 4 0 > 5 -1 > 6 1 > > How does this help me? Andrew, I think you are seeing problems where there are none. Here you have pulled out a piece of code from inside one of the limma functions. You really have no need to deal with this piece of code. It does do the right thing. If the linear model computations don't make any sense to you, it would be good idea for you to talk to a statistician at Imperial College. That would be a sensible thing to do in any case to put together a complete analysis strategy for your dataset. Best wishes Gordon > > Thanks > > > > > > Gordon K Smyth wrote: > >>On Wed, July 26, 2006 8:47 pm, Andrew Mcdonagh wrote: >> >> >>>Hi Gordon, >>> >>>Thanks for getting back to me, although I don't think you answered my >>>question. On the surface it looks like another couple of posts relating >>>to this, but that's not actually the case >>> >>>I think I might not have explained myself clearly. I have read the limma >>>guide, and I can see how you calculate the t-statistics on the >>>coefficients. The example that you give is relevant to me if the >>>coefficients themselves are comparisons between protocols. Clearly from >>>my design matrix the effect is not a comparison of protocols. So the >>>three coefficients that I estimate are the effect of protocol 1, >>>protocol 2 and protocol 3. >>> >>> >> >>Use contrasts.fit to make the coefficients the comparisons you want. >> >> >> >>>Also, the question about single channel analysis relates to this. Do you >>>have any suggestions >>> >>>Thanks >>> >>>Andy >>> >>> >> >>Please send questions to the Bioconductor list. >> >>Best wishes >>Gordon