Dear BioC List Members: I have a data set that I would like to analyze with the limma package. I am having trouble figuring out how to make the design and contrasts matrices, and I was hoping that someone would advise me. The data set contains the results from 18 Affymetrix hybridizations. The table below explains the experimental design: Sample RNA aliquot Labeled sample aliquot Dose of Tx1 +/- Tx2 Untreated - A 1 1 0 - Untreated - B 1 2 0 - Untreated - C 2 N/A 0 - Tx1 Dose1 - A 1 1 Dose1 - Tx1 Dose1 - B 1 2 Dose1 - Tx1 Dose1 - C 2 N/A Dose1 - Tx1 Dose2 - A 1 1 Dose2 - Tx1 Dose2 - B 1 2 Dose2 - Tx1 Dose2 - C 2 N/A Dose2 - Tx2 - A 1 1 0 + Tx2 - B 1 2 0 + Tx2 - C 2 N/A 0 + Tx2 + Tx1 Dose1 - A 1 1 Dose1 + Tx2 + Tx1 Dose1 - B 1 2 Dose1 + Tx2 + Tx1 Dose1 - C 2 N/A Dose1 + Tx2 + Tx1 Dose2 - A 1 1 Dose2 + Tx2 + Tx1 Dose2 - B 1 2 Dose2 + Tx2 + Tx1 Dose2 - C 2 N/A Dose2 + Tx = treatment N/A = not applicable Note that two types of technical replicates were performed: samples labeled with "A" and "B" are replicates at the level of the labeled sample aliquot, and samples labeled with "C" are replicates that were performed at the level of the RNA aliquot. In addition, the "C" replicates were done at a much later date using a different lot of microarrays, a different lot of reagents, and a different scanner. It is not surprising that we are observing a batch effect in the "C" replicates that is not removed even after normalization. Specific questions that I am hoping to have answered: 1. Can I use limma to remove the batch effect that I have observed with the third replicate? 2. If so, could someone please help me with the creation of the design and contrasts matrices? The comparisons that I would like to make are the following: Tx1 Dose1 vs. Untreated Tx1 Dose2 vs. Untreated Tx2 vs. Untreated Tx2 + Tx1 Dose1 vs. Untreated Tx2 + Tx1 Dose2 vs. Untreated 3. If anyone has any other suggestions for methods other than limma that I could use to analyze this data set, I would greatly appreciate hearing those as well. Thank you, Jim _____________________ Jim Breaux, Ph.D. ViaLogy Corp. 2400 Lincoln Ave. Altadena, CA 91001 Office: (626) 296-6473 jim.breaux@vialogy.com

This experiment has a deep problem in that there is no replication which replicates the RNA extraction process. Your replicates are just aliquots of a single RNA sample or, even worse, aliquots of a labelled sample. Hence there is no way to realistically estimate the array to array variation against which treatment effects can be compared. By comparison, the batch effect for C is not so important. The experiment is balanced with respect to A, B and C and the treatment differences, so any difference between C and the other two replicates will cancel out. I think that your only option is to rank genes and forget about format hypothesis testing. I guess that if I was analysing this data I would treat the labelled aliquots (A&B) as correlated blocks and might include a fixed effect for the C vs A,B batch effect. The block effect would be included using duplicateCorrelation() with block=c(1,1,2,3,3,4,5,5,6 etc). However this analysis will over-state the true level of significance, by an unknown amount, because of the replication problems already mentioned. The ranking of the genes will be correct if (i) the measurement error dominates or (ii) the higher level error components are proportion to the lower. The section on "Special Designs" in the limma User's Guide at http://bioinf.wehi.edu.au/limma/usersguide.pdf might be helpful. You do not make it clear what documentation you have read or whether you have already tried any possibilities. Gordon ----------- original message ------------------ Fri Oct 22 06:51:08 CEST 2004 ---------------------------------------------------------------------- ---------- Dear BioC List Members: I have a data set that I would like to analyze with the limma package. I am having trouble figuring out how to make the design and contrasts matrices, and I was hoping that someone would advise me. The data set contains the results from 18 Affymetrix hybridizations. The table below explains the experimental design: Sample RNA aliquot Labeled sample aliquot Dose of Tx1 +/- Tx2 Untreated - A 1 1 0 - Untreated - B 1 2 0 - Untreated - C 2 N/A 0 - Tx1 Dose1 - A 1 1 Dose1 - Tx1 Dose1 - B 1 2 Dose1 - Tx1 Dose1 - C 2 N/A Dose1 - Tx1 Dose2 - A 1 1 Dose2 - Tx1 Dose2 - B 1 2 Dose2 - Tx1 Dose2 - C 2 N/A Dose2 - Tx2 - A 1 1 0 + Tx2 - B 1 2 0 + Tx2 - C 2 N/A 0 + Tx2 + Tx1 Dose1 - A 1 1 Dose1 + Tx2 + Tx1 Dose1 - B 1 2 Dose1 + Tx2 + Tx1 Dose1 - C 2 N/A Dose1 + Tx2 + Tx1 Dose2 - A 1 1 Dose2 + Tx2 + Tx1 Dose2 - B 1 2 Dose2 + Tx2 + Tx1 Dose2 - C 2 N/A Dose2 + Tx = treatment N/A = not applicable Note that two types of technical replicates were performed: samples labeled with "A" and "B" are replicates at the level of the labeled sample aliquot, and samples labeled with "C" are replicates that were performed at the level of the RNA aliquot. In addition, the "C" replicates were done at a much later date using a different lot of microarrays, a different lot of reagents, and a different scanner. It is not surprising that we are observing a batch effect in the "C" replicates that is not removed even after normalization. Specific questions that I am hoping to have answered: 1. Can I use limma to remove the batch effect that I have observed with the third replicate? 2. If so, could someone please help me with the creation of the design and contrasts matrices? The comparisons that I would like to make are the following: Tx1 Dose1 vs. Untreated Tx1 Dose2 vs. Untreated Tx2 vs. Untreated Tx2 + Tx1 Dose1 vs. Untreated Tx2 + Tx1 Dose2 vs. Untreated 3. If anyone has any other suggestions for methods other than limma that I could use to analyze this data set, I would greatly appreciate hearing those as well. Thank you, Jim _____________________ Jim Breaux, Ph.D. ViaLogy Corp. 2400 Lincoln Ave. Altadena, CA 91001 Office: (626) 296-6473 jim.breaux at vialogy.com