SEQanswers grouped vs pairwise comparison for DE analysis
 Register FAQ Members List Calendar Search Today's Posts Mark Forums Read

 Similar Threads Thread Thread Starter Forum Replies Last Post ce.log Bioinformatics 17 01-13-2014 11:35 PM MenzZana Bioinformatics 2 02-04-2013 12:41 AM maembe Bioinformatics 2 06-22-2012 12:16 PM DrD2009 Bioinformatics 0 03-10-2010 12:40 PM

 06-18-2013, 07:00 AM #1 NitaC Member   Location: Philadelphia Join Date: Apr 2013 Posts: 17 grouped vs pairwise comparison for DE analysis I hope this really doesn't come across as a dumb question (newbie alert) but I have a question that has really been bugging me. To make a long story short, we determined differentially expressed genes using Cuffdiff. The conclusion for one of the comparisons doesn't exactly jive with what's accepted so we keep revisiting the issue. The terms "grouped" and "pairwise" keep coming up. Our in-house statistician prefers the grouped results but other colleagues say to use pairwise. Basically we have 8 eyes. And we're doing some comparisons between tissues. So for the grouped comparison, we take all the read counts for condition 1 in all eyes and compare it to condition 2 in all the eyes. Now, the suggestion is to compare condition 1 to condition 2 in each eye separately. And THEN somehow find a combined p-value. Another team member is using Cuffdiff and apparently there is an easy way to do this. I, however, have been using NOISeq, edgeR, and DESeq. I had grown particularly fond of NOISeq. However, to me, it just doesn't make sense to do these individual pairwise comparisons. At least for NOISeq, which performs better with biological replicates. Am I understanding something incorrectly??? Someone please clarify this for me. What is the benefit of doing the comparisons individually? To me, it's just a lot of noise because now you're getting differences between individuals.
 06-18-2013, 07:25 AM #2 dietmar13 Senior Member   Location: Vienna Join Date: Mar 2010 Posts: 107 logic question, easy answer: do NOT use cuffdiff. as i understand you have 8 eyes with each 2 tissues (=conditions). use mapper + htseq-count and R-Packages: -DESeq2 -edgeR -limma (voom function) -ShrinkBayes -pairedBayes especially the last two are very sensitive, ... all are able to use a matched pairs design (which is always better to use, if the data ARE matched pairs, with e.g. individuum (=eye) as random effect) (in my hands NOSeq and cuffdiff had the fewest significant results)
 06-18-2013, 07:45 AM #3 NitaC Member   Location: Philadelphia Join Date: Apr 2013 Posts: 17 Hi dietmar13! Thanks for your answer. My PI and statistician would be very upset to read that first line though. lol! But seriously, my teeny voice has not been enough to persuade anyone to use a different package. Still, I push on... The data really are matched pairs (ie. samples of both tissues taken from each eye). I will try again with one of the R packages. NOISeq and cuffdiff do indeed both have the fewest significant results. However, at least NOISeq gave me something. One of the comparisons with cuffdiff resulted in 1 DE gene...every other method suggested otherwise. Anyway, thanks again for your reply! -Anita
 06-18-2013, 09:19 AM #4 chadn737 Senior Member   Location: US Join Date: Jan 2009 Posts: 392 Both edgeR and DESeq/DESeq2 (not sure about NOISeq) allow you to include multiple factors into the experimental design. So you could give edgeR/DESeq/DESeq2 a table with condition as one factor and the eye/treatment as a second factor. Especially if each eye is of a different genotype or tissue. I am not fond of cuffdiff. Last edited by chadn737; 06-18-2013 at 09:21 AM.
 06-18-2013, 10:04 AM #5 NitaC Member   Location: Philadelphia Join Date: Apr 2013 Posts: 17 I am reading up on how to do this with edgeR right now. Thank you chadn737.