Dear Tom We did a direct head to head comparison of these and other microarray gene selection approach a few years back. Briefly, if you have very low sample size or high noise in the data, then Rank Products does well as its difficult to estimate the true mean or variance. Limma does better than classical statistical methods as it uses a moderated variance estimate. However as the data improves (greater sample size, higher signal:noise), classical statistical tests that utilize both mean and variance estimate do better than Rank Products. Limma also performs well in this case. Overall, we recommended limma as it performs well in across each scenario. Jeffery IB, Higgins DG, Culhane AC. (2006) Comparison and evaluation of methods for generating differentially expressed gene lists from microarray data. BMC Bioinformatics. 7:359 http://www.biomedcentral.com.ezp-prod1.hul.harvard.edu/1471-2105/7/359 Regards Aedin > Message: 1 > Date: Wed, 17 Feb 2010 15:10:49 +0100 > From: Juan Carlos Oliveros <oliveros@cnb.csic.es> > To: bioconductor@stat.math.ethz.ch > Subject: [BioC] limma and Rank Products: comparison of the number of > results > Message-ID: <4B7BF8E9.5050105@cnb.csic.es> > Content-Type: text/plain; charset=ISO-8859-1; format=flowed > > Dear all > > When working with comparative experiment based on Affymetrix gene > expression arrays I usually apply one of the following combination of > methods: > > RMA + limma + FDR > > or > > RMA+ Rank Products > (rank products "Percentage of false prediction" values are supposed to > be equivalent to FDR) > > Usually we obtain much more differentially expressed genes when using > Rank Products than when using limma at the same FDR threshold. > > I wonder if in your case is the same. Do you obtain many more results > with Rank Products than with limma at the same FDR cutoff? > > In a recent experiment we obtained the opposite (more results with > limma) and I'd like to know your experience when using both methods > regarding the number of results. > > best, > > Juan Carlos Oliveros, Ph.D. > CNB-CSIC, Madrid, Spain > > > > > > -- Aedi'n Culhane, Computational Biology and Functional Genomics Harvard School of Public Health, Dana-Farber Cancer Institute 44 Binney Street, SM822C Department of Biostatistics and Computational Biology, Dana-Farber Cancer Institute Boston, MA 02115 USA Phone: +1 (617) 632 2468 Fax: +1 (617) 582 7760 Email: aedin@jimmy.harvard.edu Web URL: http://www.hsph.harvard.edu/research/aedin-culhane/ [[alternative HTML version deleted]]

Hello Aiden, A quick question... In your fine paper, you mention that the Empirical Bayes Statistic returns the log odds that a gene is differentially expressed. This says to me that eBayes returns the probability of the alternative hypothesis given the data. Based on the F statistic, I would have thought eBayes() returns a p value of sorts, i.e., the probability of the data given the null. If eBayes really gives the probability of differential expression, that would be great, but that would require an a priori knowledge of differential expression, the very thing we are investigating with the array. Best, Tom On Feb 18, 2010, at 10:33 AM, Aedin Culhane wrote: > Dear Tom > We did a direct head to head comparison of these and other microarray > gene selection approach a few years back. > > Briefly, if you have very low sample size or high noise in the data, > then Rank Products does well as its difficult to estimate the true > mean > or variance. Limma does better than classical statistical methods as > it > uses a moderated variance estimate. > > However as the data improves (greater sample size, higher > signal:noise), > classical statistical tests that utilize both mean and variance > estimate > do better than Rank Products. Limma also performs well in this case. > > Overall, we recommended limma as it performs well in across each > scenario. > > Jeffery IB, Higgins DG, Culhane AC. (2006) Comparison and evaluation > of > methods for generating differentially expressed gene lists from > microarray data. BMC Bioinformatics. 7:359 > > http://www.biomedcentral.com.ezp- prod1.hul.harvard.edu/1471-2105/7/359 > > Regards > Aedin > >> Message: 1 >> Date: Wed, 17 Feb 2010 15:10:49 +0100 >> From: Juan Carlos Oliveros <oliveros at="" cnb.csic.es=""> >> To: bioconductor at stat.math.ethz.ch >> Subject: [BioC] limma and Rank Products: comparison of the number of >> results >> Message-ID: <4B7BF8E9.5050105 at cnb.csic.es> >> Content-Type: text/plain; charset=ISO-8859-1; format=flowed >> >> Dear all >> >> When working with comparative experiment based on Affymetrix gene >> expression arrays I usually apply one of the following combination of >> methods: >> >> RMA + limma + FDR >> >> or >> >> RMA+ Rank Products >> (rank products "Percentage of false prediction" values are supposed >> to >> be equivalent to FDR) >> >> Usually we obtain much more differentially expressed genes when using >> Rank Products than when using limma at the same FDR threshold. >> >> I wonder if in your case is the same. Do you obtain many more results >> with Rank Products than with limma at the same FDR cutoff? >> >> In a recent experiment we obtained the opposite (more results with >> limma) and I'd like to know your experience when using both methods >> regarding the number of results. >> >> best, >> >> Juan Carlos Oliveros, Ph.D. >> CNB-CSIC, Madrid, Spain >> >> >> >> >> >> > > > -- > Aedi'n Culhane, > Computational Biology and Functional Genomics > Harvard School of Public Health, Dana-Farber Cancer Institute > > 44 Binney Street, SM822C > Department of Biostatistics and Computational Biology, > Dana-Farber Cancer Institute > Boston, MA 02115 > USA > > Phone: +1 (617) 632 2468 > Fax: +1 (617) 582 7760 > Email: aedin at jimmy.harvard.edu > Web URL: http://www.hsph.harvard.edu/research/aedin-culhane/ > > > > [[alternative HTML version deleted]] > > _______________________________________________ > 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