hi Yoong, Let's continue our discussion on the Bioconductor list, as it might be useful for others as well. On Sat, Nov 23, 2013 at 10:42 AM, FeiYian Yoong <fyoong@ucdavis.edu> wrote: > Dear Mike, > > I have a couple questions about fitType on DESeq2 (v 1.2.5). I was reading > your responses on SEqAnswers. But, since I couldn't view the attached > files, I couldn't get a good idea what you meant by which fitType to use. I > have the same warning message when I ran DESeq(dds) command with my data: > > Warning messages:1: In log(ifelse(y == 0, 1, y/mu)) : NaNs produced2: step size truncated due to divergence 3: In estimateDispersionsFit(object, fitType = fitType, quiet = quiet) : > parametric fit failed, trying local fit. use plotDispEsts() to check the quality of the local fit > > The analysis was based on 1,000 genes (a subset of my 57,000 genes) as a > test. I am trying to fit these genes into a negative binomial GLM before > passing them through your independent filtering using res(). Because of the > warning message, I managed to run my data using only fitType=local. Hence, > I can't compare my fitType=local result with fitType=parametric. Also, I > ran another DESeq(dds) task using fitType=mean. My questions will be: > > 1) Is fitType=local good enough for my analysis? How can I make sure it's > as comparable as parametric? > ​The plots of dispersion you sent me have a slightly increasing slope over the mean counts​. It makes sense that the parametric fit failed, because the parametrization is y = a/x + b, a,b > 0 then dy/dx = -a/x^2 < 0 The parametrization tries to match dispersions with a decreasing slope over the mean counts. ​The local fit seems better for your data., so you can run DESeq() with fitType="local" instead. The instruction to check the quality of the fit is to make sure that the curve is not being overly influenced by individual genes, but this does not seem the case in the plot you sent me, so I would use fitType="local" > 2) Based on the graphs, would you recommend me using fitType=mean instead > of local? > > I don't really understand what it means by numerical integration on VST for > local fitType. Do you care to explain? > > This is a more technical detail about the implementation in the man page, so maybe not interesting for non-statisticians. ​The variance stabilizing transformation involves integration of the variance as a function of the mean. When the local fit is used, the integration is not a symbolic integration using the parametrized function, but a numeric integration of the local regression curve. ​ ​Mike​ [[alternative HTML version deleted]]