Dear all I am analysing data obtained from an experiment with an interwoven loop design using limma. The design and the code are listed below. Many of our probes show extremely low adjusted p-values with values low as 1.748434e-71. Hence, I was wondering whether my code somehow treats technical replication as independent ones, or whether such low p-values could be genuine. Has anyone any ideas? Many thanks for your suggestions! Pie My experimental design: We have 3 groups, A, B and C with 5 biological (independent) replicates for each group (15 RNA targets in total). The RNA's were co-hybridised to a two colour array whereby each target was twice labelled with Cy3 and twice with Cy5 in the following way: File Cy3 Cy5 File1 A1 C2 File2 A1 B1 File3 A2 C3 File4 A2 B2 File5 A3 C4 File6 A3 B3 File7 A4 C5 File8 A4 B4 File9 A5 C1 File10 A5 B5 File11 B1 A3 File12 B1 C1 File13 B2 C2 File14 B2 A4 File15 B3 C3 File16 B3 A5 File17 B4 C4 File18 B4 A1 File19 B5 C5 File20 B5 A2 File21 C1 A2 File22 C1 B3 File23 C2 A3 File24 C2 B4 File25 C3 A4 File26 C3 B5 File27 C4 A5 File28 C4 B1 File29 C5 A1 File30 C5 B2 My code for fitting the linear model: design=modelMatrix(targets, ref="A1") cor=duplicateCorrelation(MA, design, ndups=4, spacing=1, weights=w) fit=lmFit(MA, cor=cor$consensus.correlation, design, ndups=4, spacing=1, weights=w) cont.matrix=makeContrasts(AvsB=(A2+A3+A4+A5-B1-B2-B3-B4-B5)/5, AvsC=(A2+A3+A4+A5-C1-C2-C3-C4-C5)/5, CvsB=(C1+C2+C3+C4+C5-B1-B2-B3-B4-B5)/5, levels=design) fit2=contrasts.fit(fit, cont.matrix) fit2=eBayes(fit2) topTable(fit2, coef="AvsB", adjust.method="fdr", sort.by="p") ------------------------------------- Dr Pie M?ller Vector Group Liverpool School of Tropical Medicine Pembroke Place Liverpool L3 5QA UK Tel +44(0) 151 705 3225 Fax +44(0) 151 705 3369 http://www.liv.ac.uk/lstm http://www.ivcc.com

Yes, your code is treating the technical replicates as if they were the biological replicates and the biological replicates as if they were different treatments. This is because A1 and A2 are each given a factor. You need to rename all of the A's with the name "A", similarly for the Bs and Cs. --Naomi At 06:13 AM 9/17/2007, Muller, Pie wrote: >Dear all > >I am analysing data obtained from an experiment >with an interwoven loop design using limma. The >design and the code are listed below. Many of >our probes show extremely low adjusted p-values >with values low as 1.748434e-71. Hence, I was >wondering whether my code somehow treats >technical replication as independent ones, or >whether such low p-values could be genuine. Has anyone any ideas? > >Many thanks for your suggestions! > >Pie > > >My experimental design: > >We have 3 groups, A, B and C with 5 biological >(independent) replicates for each group (15 RNA >targets in total). The RNA's were co-hybridised >to a two colour array whereby each target was >twice labelled with Cy3 and twice with Cy5 in the following way: > >File Cy3 Cy5 > >File1 A1 C2 >File2 A1 B1 >File3 A2 C3 >File4 A2 B2 >File5 A3 C4 >File6 A3 B3 >File7 A4 C5 >File8 A4 B4 >File9 A5 C1 >File10 A5 B5 >File11 B1 A3 >File12 B1 C1 >File13 B2 C2 >File14 B2 A4 >File15 B3 C3 >File16 B3 A5 >File17 B4 C4 >File18 B4 A1 >File19 B5 C5 >File20 B5 A2 >File21 C1 A2 >File22 C1 B3 >File23 C2 A3 >File24 C2 B4 >File25 C3 A4 >File26 C3 B5 >File27 C4 A5 >File28 C4 B1 >File29 C5 A1 >File30 C5 B2 > > >My code for fitting the linear model: > >design=modelMatrix(targets, ref="A1") >cor=duplicateCorrelation(MA, design, ndups=4, spacing=1, weights=w) >fit=lmFit(MA, cor=cor$consensus.correlation, >design, ndups=4, spacing=1, weights=w) >cont.matrix=makeContrasts(AvsB=(A2+A3+A4+A5-B1-B2-B3-B4-B5)/5, >AvsC=(A2+A3+A4+A5-C1-C2-C3-C4-C5)/5, >CvsB=(C1+C2+C3+C4+C5-B1-B2-B3-B4-B5)/5, levels=design) >fit2=contrasts.fit(fit, cont.matrix) >fit2=eBayes(fit2) >topTable(fit2, coef="AvsB", adjust.method="fdr", sort.by="p") > > >------------------------------------- > >Dr Pie M?ller >Vector Group >Liverpool School of Tropical Medicine >Pembroke Place >Liverpool >L3 5QA >UK > >Tel +44(0) 151 705 3225 >Fax +44(0) 151 705 3369 > >http://www.liv.ac.uk/lstm >http://www.ivcc.com > >_______________________________________________ >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 Naomi S. Altman 814-865-3791 (voice) Associate Professor Dept. of Statistics 814-863-7114 (fax) Penn State University 814-865-1348 (Statistics) University Park, PA 16802-2111

Pie, I do not recall all the examples but you have: 2 color arrays - hence correlation on the same array 4 technical reps - hence correlation on the same biological replicates I did not check your code before, but I think that you need to show what is in "targets" for me to help you more. I think that to handle this analysis, you need to use single channel analysis. But then you have 2 sources of dependence, and limma cannot handle this. --Naomi At 10:16 AM 9/17/2007, Muller, Pie wrote: >Naomi, > >Thank you very much for your reply, the p-values >seem to make much more sense now although I am >still slightly confused. I tried to follow >Example 8.2 of the Limma User's Guide by taking >into account that RNA from each individual >(e.g., "A1") appeared on four arrays. Why would >my previous experimental design not follow the same logic as in example 8.2? > >Apologies for coming back on this... > >Thanks, >Pie > >-----Original Message----- >From: Naomi Altman [mailto:naomi at stat.psu.edu] >Sent: 17 September 2007 14:32 >To: Muller, Pie; bioconductor at stat.math.ethz.ch >Subject: Re: [BioC] Extremely low p-values in limma > >Yes, your code is treating the technical >replicates as if they were the biological >replicates and the biological replicates as if >they were different treatments. This is because >A1 and A2 are each given a factor. You need to >rename all of the A's with the name "A", similarly for the Bs and Cs. > >--Naomi > >At 06:13 AM 9/17/2007, Muller, Pie wrote: > >Dear all > > > >I am analysing data obtained from an experiment > >with an interwoven loop design using limma. The > >design and the code are listed below. Many of > >our probes show extremely low adjusted p-values > >with values low as 1.748434e-71. Hence, I was > >wondering whether my code somehow treats > >technical replication as independent ones, or > >whether such low p-values could be genuine. Has anyone any ideas? > > > >Many thanks for your suggestions! > > > >Pie > > > > > >My experimental design: > > > >We have 3 groups, A, B and C with 5 biological > >(independent) replicates for each group (15 RNA > >targets in total). The RNA's were co-hybridised > >to a two colour array whereby each target was > >twice labelled with Cy3 and twice with Cy5 in the following way: > > > >File Cy3 Cy5 > > > >File1 A1 C2 > >File2 A1 B1 > >File3 A2 C3 > >File4 A2 B2 > >File5 A3 C4 > >File6 A3 B3 > >File7 A4 C5 > >File8 A4 B4 > >File9 A5 C1 > >File10 A5 B5 > >File11 B1 A3 > >File12 B1 C1 > >File13 B2 C2 > >File14 B2 A4 > >File15 B3 C3 > >File16 B3 A5 > >File17 B4 C4 > >File18 B4 A1 > >File19 B5 C5 > >File20 B5 A2 > >File21 C1 A2 > >File22 C1 B3 > >File23 C2 A3 > >File24 C2 B4 > >File25 C3 A4 > >File26 C3 B5 > >File27 C4 A5 > >File28 C4 B1 > >File29 C5 A1 > >File30 C5 B2 > > > > > >My code for fitting the linear model: > > > >design=modelMatrix(targets, ref="A1") > >cor=duplicateCorrelation(MA, design, ndups=4, spacing=1, weights=w) > >fit=lmFit(MA, cor=cor$consensus.correlation, > >design, ndups=4, spacing=1, weights=w) > >cont.matrix=makeContrasts(AvsB=(A2+A3+A4+A5-B1-B2-B3-B4-B5)/5, > >AvsC=(A2+A3+A4+A5-C1-C2-C3-C4-C5)/5, > >CvsB=(C1+C2+C3+C4+C5-B1-B2-B3-B4-B5)/5, levels=design) > >fit2=contrasts.fit(fit, cont.matrix) > >fit2=eBayes(fit2) > >topTable(fit2, coef="AvsB", adjust.method="fdr", sort.by="p") > > > > > >------------------------------------- > > > >Dr Pie M?ller > >Vector Group > >Liverpool School of Tropical Medicine > >Pembroke Place > >Liverpool > >L3 5QA > >UK > > > >Tel +44(0) 151 705 3225 > >Fax +44(0) 151 705 3369 > > > >http://www.liv.ac.uk/lstm > >http://www.ivcc.com > > > >_______________________________________________ > >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 > >Naomi S. Altman 814-865-3791 (voice) >Associate Professor >Dept. of Statistics 814-863-7114 (fax) >Penn State University 814-865-1348 (Statistics) >University Park, PA 16802-2111 > >_______________________________________________ >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 Naomi S. Altman 814-865-3791 (voice) Associate Professor Dept. of Statistics 814-863-7114 (fax) Penn State University 814-865-1348 (Statistics) University Park, PA 16802-2111

I have not used MAANOVA, but according to what I have read, it will handle multiple random effects, which in your case would be array and biological sample. --Naomi At 05:28 AM 9/18/2007, Muller, Pie wrote: >Dear Naomi, > >I have attached a graphical representation of >our experimental design in a jpg image to give >you a quick idea. In our experiments the "targets" look as follows: > >File Cy3 Cy5 > >File1 A1 C2 >File2 A1 B1 >File3 A2 C3 >File4 A2 B2 >File5 A3 C4 >File6 A3 B3 >File7 A4 C5 >File8 A4 B4 >File9 A5 C1 >File10 A5 B5 >File11 B1 A3 >File12 B1 C1 >File13 B2 C2 >File14 B2 A4 >File15 B3 C3 >File16 B3 A5 >File17 B4 C4 >File18 B4 A1 >File19 B5 C5 >File20 B5 A2 >File21 C1 A2 >File22 C1 B3 >File23 C2 A3 >File24 C2 B4 >File25 C3 A4 >File26 C3 B5 >File27 C4 A5 >File28 C4 B1 >File29 C5 A1 >File30 C5 B2 > > >If limma cannot handle the above structure would >"maanova" be a suitable package for the analysis of this type of experiment? > >Best, >Pie > > > > >-----Original Message----- >From: Naomi Altman [mailto:naomi at stat.psu.edu] >Sent: 17 September 2007 20:44 >To: Muller, Pie; Naomi Altman; bioconductor at stat.math.ethz.ch >Subject: Re: [BioC] Extremely low p-values in limma > >Pie, > >I do not recall all the examples but you have: > >2 color arrays - hence correlation on the same array >4 technical reps - hence correlation on the same biological replicates > >I did not check your code before, but I think >that you need to show what is in "targets" for me to help you more. > >I think that to handle this analysis, you need to >use single channel analysis. But then you have 2 >sources of dependence, and limma cannot handle this. > >--Naomi > >At 10:16 AM 9/17/2007, Muller, Pie wrote: > >Naomi, > > > >Thank you very much for your reply, the p-values > >seem to make much more sense now although I am > >still slightly confused. I tried to follow > >Example 8.2 of the Limma User's Guide by taking > >into account that RNA from each individual > >(e.g., "A1") appeared on four arrays. Why would > >my previous experimental design not follow the same logic as in example 8.2? > > > >Apologies for coming back on this... > > > >Thanks, > >Pie > > > >-----Original Message----- > >From: Naomi Altman [mailto:naomi at stat.psu.edu] > >Sent: 17 September 2007 14:32 > >To: Muller, Pie; bioconductor at stat.math.ethz.ch > >Subject: Re: [BioC] Extremely low p-values in limma > > > >Yes, your code is treating the technical > >replicates as if they were the biological > >replicates and the biological replicates as if > >they were different treatments. This is because > >A1 and A2 are each given a factor. You need to > >rename all of the A's with the name "A", similarly for the Bs and Cs. > > > >--Naomi > > > >At 06:13 AM 9/17/2007, Muller, Pie wrote: > > >Dear all > > > > > >I am analysing data obtained from an experiment > > >with an interwoven loop design using limma. The > > >design and the code are listed below. Many of > > >our probes show extremely low adjusted p-values > > >with values low as 1.748434e-71. Hence, I was > > >wondering whether my code somehow treats > > >technical replication as independent ones, or > > >whether such low p-values could be genuine. Has anyone any ideas? > > > > > >Many thanks for your suggestions! > > > > > >Pie > > > > > > > > >My experimental design: > > > > > >We have 3 groups, A, B and C with 5 biological > > >(independent) replicates for each group (15 RNA > > >targets in total). The RNA's were co-hybridised > > >to a two colour array whereby each target was > > >twice labelled with Cy3 and twice with Cy5 in the following way: > > > > > >File Cy3 Cy5 > > > > > >File1 A1 C2 > > >File2 A1 B1 > > >File3 A2 C3 > > >File4 A2 B2 > > >File5 A3 C4 > > >File6 A3 B3 > > >File7 A4 C5 > > >File8 A4 B4 > > >File9 A5 C1 > > >File10 A5 B5 > > >File11 B1 A3 > > >File12 B1 C1 > > >File13 B2 C2 > > >File14 B2 A4 > > >File15 B3 C3 > > >File16 B3 A5 > > >File17 B4 C4 > > >File18 B4 A1 > > >File19 B5 C5 > > >File20 B5 A2 > > >File21 C1 A2 > > >File22 C1 B3 > > >File23 C2 A3 > > >File24 C2 B4 > > >File25 C3 A4 > > >File26 C3 B5 > > >File27 C4 A5 > > >File28 C4 B1 > > >File29 C5 A1 > > >File30 C5 B2 > > > > > > > > >My code for fitting the linear model: > > > > > >design=modelMatrix(targets, ref="A1") > > >cor=duplicateCorrelation(MA, design, ndups=4, spacing=1, weights=w) > > >fit=lmFit(MA, cor=cor$consensus.correlation, > > >design, ndups=4, spacing=1, weights=w) > > >cont.matrix=makeContrasts(AvsB=(A2+A3+A4+A5-B1-B2-B3-B4-B5)/5, > > >AvsC=(A2+A3+A4+A5-C1-C2-C3-C4-C5)/5, > > >CvsB=(C1+C2+C3+C4+C5-B1-B2-B3-B4-B5)/5, levels=design) > > >fit2=contrasts.fit(fit, cont.matrix) > > >fit2=eBayes(fit2) > > >topTable(fit2, coef="AvsB", adjust.method="fdr", sort.by="p") > > > > > > > > >------------------------------------- > > > > > >Dr Pie M?ller > > >Vector Group > > >Liverpool School of Tropical Medicine > > >Pembroke Place > > >Liverpool > > >L3 5QA > > >UK > > > > > >Tel +44(0) 151 705 3225 > > >Fax +44(0) 151 705 3369 > > > > > >http://www.liv.ac.uk/lstm > > >http://www.ivcc.com > > > > > >_______________________________________________ > > >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 > > > >Naomi S. Altman 814-865-3791 (voice) > >Associate Professor > >Dept. of Statistics 814-863-7114 (fax) > >Penn State University 814-865-1348 (Statistics) > >University Park, PA 16802-2111 > > > >_______________________________________________ > >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 > >Naomi S. Altman 814-865-3791 (voice) >Associate Professor >Dept. of Statistics 814-863-7114 (fax) >Penn State University 814-865-1348 (Statistics) >University Park, PA 16802-2111 > Naomi S. Altman 814-865-3791 (voice) Associate Professor Dept. of Statistics 814-863-7114 (fax) Penn State University 814-865-1348 (Statistics) University Park, PA 16802-2111