Hi Ina, I don't think voom has any special consideration for observation outliers, but limma's 'eBayes' function has a 'robust' argument which I believe has the same effect as the corresponding argument in edgeR's 'estimateDisp', i.e. dealing with outlier tags that have abnormally high (or low) variance. -Ryan On Fri 21 Mar 2014 08:48:23 AM PDT, Ina Hoeschele wrote: > Hi Mark, > how would the presence of observation outliers potentially causing dispersion outliers be handled in voom? > Many thanks, Ina > > > ----- Original Message ----- > From: "Mark Robinson" <mark.robinson at="" imls.uzh.ch=""> > To: "Davide Cittaro" <cittaro.davide at="" hsr.it=""> > Cc: "Bioconductor mailing list" <bioconductor at="" r-project.org="">, "xiaobei.zhou at uzh.ch Zhou" <xiaobei.zhou at="" uzh.ch="">, "Gordon Smyth" <smyth at="" wehi.edu.au=""> > Sent: Thursday, March 20, 2014 9:58:05 AM > Subject: Re: [BioC] tagwise parameters for negative binomial distribution in edgeR > > Hi Davide, > > Just to add another option, there is also estimateGLMRobustDisp(), which is a wrapper for an iteratively re-weighted scheme that down weights outliers -- effectively using a constant prior degrees of freedom. This was developed independently of estimateDisp() from a different perspective. But, what this gives is a matrix of weights (so, one for each observation, not just for each tag) where identified outliers should exhibit lower weights. So, you could use this to identify outliers observation-wise or tag-wise (e.g. take column sum of weights). You'd want >= 3 replicates per condition for this one though. > > In code, you could do something like: > > d <- estimateGLMRobustDisp(d, design) > summary(d$weights) > > More details can be found at: > > http://arxiv.org/abs/1312.3382 > > http://imlspenticton.uzh.ch/robinson_lab/edgeR_robust/supplement.pdf > (in particular, Supplementary Figure 8, which shows ROC curves for ability to separate [simulated] outliers by weights/residuals and yet another option is DESeq2's Cook's [observation-wise or max-by-tag] distance; we don't have a curve for estimateDisp!) > > > Best regards, Mark > > > ---------- > Prof. Dr. Mark Robinson > Bioinformatics, Institute of Molecular Life Sciences > University of Zurich > http://ow.ly/riRea > > > > > > > On 20.03.2014, at 01:04, Gordon K Smyth <smyth at="" wehi.edu.au=""> wrote: > >> Dear Davide, >> >> Do you want to identify tags (genes) with dispersion values that are so high (relative to other genes with similar count sizes) that they should be considered outliers? >> >> The easiest way to do this is to use >> >> d <- estimateDisp(d, design, robust=TRUE) >> >> and then look at the output values for prior.df: >> >> summary(d$prior.df) >> >> Any tag with a small prior.df is considered an outlier. You can sort tags by their prior.df values to select the most significant outliers. >> >> Note that the methodology used by the estimateDisp() robust procedure is more complicated than simply using NB probabilities, because one has to take into acccount the genome-wide distribution of the dispersion values as well as accounting for the fact that the fitted values (p) have been estimated from the same data. The methodology is mostly explained in: >> >> http://www.statsci.org/smyth/pubs/edgeRChapterPreprint.pdf >> http://www.statsci.org/smyth/pubs/RobustEBayesPreprint.pdf >> >> Best wishes >> Gordon >> >>> From: Davide Cittaro <cittaro.davide at="" hsr.it=""> >>> To: "bioconductor at r-project.org list" <bioconductor at="" r-project.org=""> >>> Subject: [BioC] tagwise parameters for negative binomial distribution >>> in edgeR >>> >>> Dear list, >> >>> I have a DGElist object in edgeR, already processed with calcNormFactors, estimateCommonDispersion and estimateTagWiseDispersion. Now, I would like to identify tagwise outliers in my data, I thought I could estimate NB distribution for each tag. Given that a NB is defined by two parameters (r and p), I assume that r = 1/x$tagwise.dispersion, how can I get tagwise p from DGEList dataframe? >> >>> Thanks >>> >>> d >> >> ______________________________________________________________________ >> The information in this email is confidential and intend...{{dropped:4}} >> >> _______________________________________________ >> 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 > > _______________________________________________ > 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 > > _______________________________________________ > 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

On 21/mar/2014, at 01:47, Gordon K Smyth <smyth at="" wehi.edu.au=""> wrote: >> >> Mmm, actually I would like to identify the sample that is an outlier for >> a specific gene, that's why I thought I could focus on tagwise >> distribution. > > See Mark Robinson's post. > > It depends on your purpose however. Do you want to downweight/ignore > outliers, or do you want to identify them because they are interesting? In this case outliers may be relevant, especially the less represented. I'm running the estimateGLMRobustDisp approach (although it takes a loong time) >>> Any tag with a small prior.df is considered an outlier. You can sort tags >>> by their prior.df values to select the most significant outliers. >> >> Does this identify a tag that is an outlier over all samples? > > Basically yes. We distinguish dispersion outliers and observation > outliers. An observation outlier is an individual count that is an > outlier (relative to other counts for the same gene). A dispersion > outlier is a gene that shows much more variability between replicates than > other genes at the same cpm level. A dispersion outlier may arise from > one or more observation outliers, but not necessarily. It could also > arise from systematically larger variability. Thanks for the explanation. Best, d