http://arxiv.org/pdf/1301.5277v2.pdf may be of interest voom() transforms everything to log2(RPM) before estimating mean- variance trends, so fractional count estimates from RSEM won't impact the modeling anyhow. See also http://www.statsci.org/smyth/pubs/VoomPreprint.pdf Personally I've wondered for a long, long time why people spend so much time on GLMs for sensitivity, when more biological replicates would make more difference. On Mon, Aug 19, 2013 at 1:44 PM, Steve Lianoglou <lianoglou.steve@gene.com>wrote: > Hi, > > On Mon, Aug 19, 2013 at 11:17 AM, Alan Smith <alan.sm310@gmail.com> wrote: > > Hello, > > > > I used RSEM to extract gene and transcript level read count information > for > > our single end read libraries. Then rounded off the expected read counts > to > > use for differential expression analysis using edgeR at both transcript > and > > gene level. However, I found that the number of DE transcripts were > almost > > 10 times less than those of genes. > > > > Is this expected or should I be following other package to analyze > > transcript level DE analysis. > > If you take a minute to take a walk down memory lane and browse > through the list archives searching for "RSEM": > > > http://search.gmane.org/?query=rsem&author=&group=gmane.science.biol ogy.informatics.conductor&sort=date&DEFAULTOP=and&xP=Zrsem&xFILTERS=Gs cience.biology.informatics.conductor---A/ > > You'll find that RSEM output doesn't play well with edgeR and DESeq. > These methods explicitly require *count* data -- not any old number > that has been then rounded to an integer. > > I recall a thread about how voom would likely work well with RSEM > output, but I can't seem to dig it up right now -- I'm only finding > other people mentioning that thread ;-) > > HTH, > -steve > > -- > Steve Lianoglou > Computational Biologist > Bioinformatics and Computational Biology > Genentech > > _______________________________________________ > Bioconductor mailing list > Bioconductor@r-project.org > https://stat.ethz.ch/mailman/listinfo/bioconductor > Search the archives: > http://news.gmane.org/gmane.science.biology.informatics.conductor > -- *He that would live in peace and at ease, * *Must not speak all he knows, nor judge all he sees.* * * Benjamin Franklin, Poor Richard's Almanack<http: archive.org="" details="" poorrichardsalma00franrich=""> [[alternative HTML version deleted]]

On Mon, Aug 19, 2013 at 4:44 PM, Steve Lianoglou <lianoglou.steve@gene.com>wrote: > You'll find that RSEM output doesn't play well with edgeR and DESeq. > These methods explicitly require *count* data -- not any old number > that has been then rounded to an integer. > I've never actually understood/bought-into this line of reasoning. RSEM values are not "any old number", but are just a special form of counts; if you sum all the values, you get the total mapped read count. If you'd like to split hairs, they are weighted counts that are represented succinctly (perhaps even sufficiently?) by multiplying each count by its weight, rather than supplying the individual raw counts and their weights as separate vectors. I don't think anyone would argue against our numerical ability to include weighted data in any GLM, so why so much fuss? Since edgeR/DESeq/voom/etc do not themselves handle the ambiguity of read mapping (cf. any Cufflinks paper decrying this fact), it seems like RSEM + edgeR/DESeq/voom/etc is your only analytical option for ANOVA-type hypothesis testing that attempts to take read mapping uncertainty into account (albeit without integrating that uncertainty in the model itself, as cuffdiff/eXpress do). This all seems squarely in the realm of "all models are wrong, but some are useful" -- we're just making choices about what kind of "wrongness" we're more willing to tolerate (rounding RSEM weighted counts: perhaps distributional wrongness; taking raw read counts with no uncertainty: certain error/variance wrongness). -Aaron [[alternative HTML version deleted]]