Hello, I am trying to use DESeq to test for differential expression in a multi-factor, multi-condition experiment. Factor 1 (Blocking) has 2 conditions, Factor 2 (Time) has 4 conditions and Factor 3 (exposure) has 2 conditions. I am interested in making pairwise comparisons between time points while utilizing a model containing Blocking and Exposure parameters: fit1 <- nbinomFitGLM(cds, count ~ Blocking + Exposure ) fit2 <- nbinomFitGLM(cds, count ~ Blocking + Exposure + Time) pvals <- nbinomGLMTest( fit2, fit1 ) I am unclear whether the significant genes returned by this analysis are significant across all time points or any time point. Is there a way to test pairwise (time1 vs. time2) while preserving the blocking and exposure components (ie not using nbinomTest to directly compare time1 and time 2) Thanks for the help! Andrew -- Andrew Lysaght Cyclocross Captain, MIT Cycling Ph.D. Candidate Program in Speech and Hearing Bioscience and Technology Harvard/MIT Division of Health Sciences and Technology [[alternative HTML version deleted]]

Hi Andrew On 15/03/13 22:56, Andrew Lysaght wrote: [...] > fit1 <- nbinomFitGLM(cds, count ~ Blocking + Exposure ) > fit2 <- nbinomFitGLM(cds, count ~ Blocking + Exposure + Time) > pvals <- nbinomGLMTest( fit2, fit1 ) > > I am unclear whether the significant genes returned by this analysis are > significant across all time points or any time point. Is there a way to > test pairwise (time1 vs. time2) while preserving the blocking and exposure > components (ie not using nbinomTest to directly compare time1 and time 2) Significance with the test you use above means that you can reject the null hypothesis that expression does not change over time, i.e. that at least one time point is different. To compare two specific time points, subset the cds object to the samples from these two time point before the test: cds2 <- cds[ , pData(cds)$Time %in% c( "timepoint1", "timepoint2" ) ] where "timepoint1" and "timepoint2" are the levels describing the two time points you want to compare. Then run nbinomFitGLMs and nbinomGLMTest as above on cds2. Unless noise (within-group variability) is very different in different time points, you get better power if you do the subsetting _after_ calling estimateDispersion. Simon