Hi, Is there an R/Bioconductor package that can do error-weighted ANOVAs for microarray data? For every sample and every transcript I have both a value estimate and an error estimate and I'd like to use the error estimates to weight the fold changes and p-values. The error-weighting model of interest is described here: http://bioinformatics.oxfordjournals.org/content/22/9/1111.full.pdf+ht ml Background: I have inherited a project that was started using Rosetta Resolver software. Rosetta no longer exists, and the Resolver software was purchased by Microsoft only to be discontinued. Microarray gene expression data were generated from Affy chips, and I have access to the data that are already normalized/pre-processed. Resolver has generated normalized data with the following 3 pieces of information for EACH sample and EACH transcript: value, p-value and error estimate (the p-value and error estimate are estimated for each sample, it's not a group estimate). I also have group-level comparison p-value and fold-change results generated from Resolver. When I run simple means and t-test/anovas on the "values" for each sample (e.g., using genefilter rowFtests) I get slightly different fold changes and slightly to significantly different p-values than what is given from Resolver. I assume this is because of the error- weighting performed as part of the one-way ANOVAs in Resolver. I'd like to recreate the results from Resolver, as a significant amount of down-stream analyses have been performed on the Resolver-derived fold change and p-value data, but am not particularly comfortable using the FC and p-value data given to me if I can't replicate it myself. Any insight would be greatly appreciated! Sincerely, Ida

Well, if you want to exactly replicate the Rosetta computations, you are probably going to have to implement them yourself. Based on my reading of the paper, you must use the equations in section 3.1 to compute the error-weighted average expression of each group along with the corresponding combined error for each group. Then use section 2.2 to compute the log-ratio and associated error (i.e. "xdev") between groups. Finally use section 3.3 to compute the p-value. As an alternative analysis method that also makes use of error estimates, you can use the lm function or limma's lmFit with the "weights" argument, using 1/variance as the value for weights. Assuming that the error estimates are provided as standard deviations, you would compute the weight as 1/error^2. Note that this is the same weighting scheme used in Section 3.1, Equation 16. Hope this helps, -Ryan On Wed 05 Mar 2014 11:21:20 AM PST, Ida Hatoum wrote: > Hi, > > Is there an R/Bioconductor package that can do error-weighted ANOVAs for microarray data? For every sample and every transcript I have both a value estimate and an error estimate and I'd like to use the error estimates to weight the fold changes and p-values. > > The error-weighting model of interest is described here: > http://bioinformatics.oxfordjournals.org/content/22/9/1111.full.pdf+ html > > Background: > I have inherited a project that was started using Rosetta Resolver software. Rosetta no longer exists, and the Resolver software was purchased by Microsoft only to be discontinued. Microarray gene expression data were generated from Affy chips, and I have access to the data that are already normalized/pre-processed. Resolver has generated normalized data with the following 3 pieces of information for EACH sample and EACH transcript: value, p-value and error estimate (the p-value and error estimate are estimated for each sample, it's not a group estimate). I also have group-level comparison p-value and fold-change results generated from Resolver. When I run simple means and t-test/anovas on the "values" for each sample (e.g., using genefilter rowFtests) I get slightly different fold changes and slightly to significantly different p-values than what is given from Resolver. I assume this is because of the error- weighting performed as part of the one-way ANOVAs in Resolve! > r. I'd like to recreate the results from Resolver, as a significant amount of down-stream analyses have been performed on the Resolver-derived fold change and p-value data, but am not particularly comfortable using the FC and p-value data given to me if I can't replicate it myself. > > Any insight would be greatly appreciated! > > Sincerely, > > Ida > _______________________________________________ > Bioconductor mailing list > Bioconductor at r-project.org > https://stat.ethz.ch/mailman/listinfo/bioconductor > Search the archives: http://news.gmane.org/gmane.science.biology.informatics.conductor