Yifang It really is not important, and there is little to be gained by looking at these values. Once normalization has been applied (or when it is necessary), you cannot draw reliable conclusions from looking at the raw counts or the pseudocounts, or the library sizes.
We use the normalization factors to adjust the library sizes so that we can compare the expression level of a gene across different samples, the composition of which may be quite different. When we allow different compositions of different samples (e.g. sets of genes that are very highly expressed in one sample but not in others) then there is absolutely no requirement or expectation for the new library sizes to be approximately equal across all samples. The DE testing is still valid---in fact even more valid than if we were not to use the normalization factors to adjust the effective library size. In the context of the mailing list I'm not sure that I can offer any more explanation. I could perhaps provide a little example to demonstrate further (although I had hoped that by working through Mark R's example that I posted previously you would get a feel for how/why TMM normalization works and is useful). However, I don't really want to distract you with more stuff that is all only tangential to your real question of interest, which I'm sure is finding out which genes in your dataset show significant evidence of differential expression. The library sizes adjusted or not are not important for this task. Cheers Davis On Jun 25, 2011, at 8:13 AM, yifang tan wrote: > Thanks Davis. > The problems are clear now. Just curious about the two formulas > > >>> norm_counts.table <- t(t(d$pseudo.alt)*(d$samples$norm.factors)) > >>> This is *not* correct. The TMM normalization factors refer to library > >>> sizes only and should not be used to normalize counts in this way. > > However, I checked the sum of each library that are VERY close to the > common.lib.size. as > # Zygote1 Zygote2 Octant1 Octant2 Globular1 Globular2 Heart1 > Heart2 Torpedo1 Torpedo2 Bent1 Bent2 Mature1 Mature2 > # 16538549 16542204 16541874 16540999 16544275 16548353 16536743 > 16539129 16541883 16546470 16545841 16547124 16544220 16541251 > > > d$common.lib.size > [1] 16554344.47 > > While using your formula: > >>> norm_cpm.table <- 1e06*t(t(d$counts) / d$samples$lib.size) > I got colSums (norm_cpm.table) as: > # Zygote1 Zygote2 Octant1 Octant2 Globular1 Globular2 Heart1 > Heart2 Torpedo1 Torpedo2 Bent1 Bent2 Mature1 Mature2 > # 999024.9 999414.7 999240.4 999097.4 999444.2 999516.7 > 998947.1 999007.6 999268.1 999407 999502.6 999476 999255.3 > 999224.1 > > Or, if I use > >>> cpm <- 1e06*t(t(d$counts) / (d$samples$lib.size*d$samples$norm.factors) ) > I got ColSums as following : > Zygote1 Zygote2 Octant1 Octant2 Globular1 Globular2 > Heart1 Heart2 Torpedo1 Torpedo2 Bent1 Bent2 > Mature1 Mature2 > 1300191.3 1307096.6 877903.3 727683.3 1142933.3 1113715.6 > 710072.8 770758.1 679471.8 689203.1 972693.7 1086163.1 > 1617750.8 1635113.6 > > The last set is similar to the one that brought my question in my first > email. The point is with the other two. The first formula gave numbers close > to common.lib.size, which made sense to me and that's why I thought it is the > right formula in my correction email, but it is wrong! The second formula > gave similar sum counts for each library, which are quite different to the > common.lib.size, but it is the right one! Maybe this is not important to me > anymore, but it still bugs me as it was not addressed in the paper (Robinson > & Oshlack, 2010) or in the manuals. Do you by chance have any explanation > about this? > Thank you! > > Yifang > > > Yifang Tan > > > > From: [email protected] > Subject: Re: [Bioc-sig-seq] edgeR common library size question. > Date: Fri, 24 Jun 2011 12:21:03 +1000 > To: [email protected] > > Hi Yifang > > I've received a few emails with edgeR questions from you overnight, two of > which look the same, so I'll try to answer all of your questions here and > hope that I haven't missed anything important. > > (1) > Normalized counts are a useful way visually to compare relative expression > levels across samples for a gene---but they should not (of course) be used > for DE analysis in edgeR directly. > > In a later email you proposed this formula for computing normalized counts: > norm_counts.table <- t(t(d$pseudo.alt)*(d$samples$norm.factors)) > > This is *not* correct. The TMM normalization factors refer to library sizes > only and should not be used to normalize counts in this way. At the moment we > favour looking at counts per million as a normalized measure---it is easy to > understand how they are calculated and the scale is a convenient one for > interpretation. As I wrote in my previous email, these can be computed as > follows: > cpm <- 1e06*t( t(d$counts) / d$samples$lib.size ) > If you wish to include TMM normalization to adjust the library sizes when > getting counts per million (a good idea), then this will do the job: > cpm <- 1e06*t( t(d$counts) / (d$samples$lib.size*d$samples$norm.factors) ) > In principle, the pseudocounts are an appropriate normalized version of the > counts for visual comparison, but as we've seen they become less transparent > once normalization factors are used in the analysis. > > The TMM normalization factors change the *effective* library size, they are > not used to normalize the *count data*. That is what is meant by > normalization in this context. We still make a library size adjustment, but > we normalize the library sizes with TMM to deal with compositional bias etc. > It really is beyond the limits of my time and desire to explain here in > greater detail the hows and whys. The TMM paper (Robinson & Oshlack, 2010) > has the full explanations, so I would refer you there. > > (2) > You are losing your rownames because of the way you are manipulating your > data in R---this is not an edgeR problem. Your code: > countsTable <- read.delim("Deep_seq_data.csv", header=T, stringsAsFactors=T) > > # Remove the first column which is FeatureID as AGI number > d<-countsTable[, -1] > > conditions<-rep(c("Zygote", "Octant", "Globular", "Heart", "Torpedo", "Bent", > "Mature"), each=2) > > #set the row name for database, believed to be mySQL > rownames(d)<-countsTable[,1] > > You assign rownames to the object "d" here. > > # make a DGE object > d<-DGEList(counts=countsTable[,-1], group=conditions) > > but when you define your DGEList, you use countsTable[,-1], which (I can only > assume) has no rownames. > > If you simply call the line rownames(d)<-countsTable[,1] *after* you have > defined your DGEList object d, then you will have all of your rownames > throughout the rest of your analysis. > > This problem: > detags.com.ZO <- rownames(topTags(de.com.ZO, n=20592)$table) #n=20592 is > the total row number in the analysis > > #Always got problem with this line even I declare the total number of rows > in full as above. > > d$counts[detags.com.ZO, ] > Error: subscript out of bounds > > Should disappear once you have your rownames correctly defined as suggested > above. I believe the problem here is that rownames(d$counts) is NULL, and > rownames(topTags(de.com.ZO, n=Inf)$table) are "tag.1", "tag.2" etc., so they > do not match up. If you want to include all genes in your data set, then the > neatest way is to set n=Inf in your call to topTags as shown here. > > Hope all of that gets you running smoothly. > > Best wishes > Davis > > > > On Jun 24, 2011, at 6:31 AM, yifang tan wrote: > > Thanks Davis! > > It went very well after I check the number with > "colSums(d$pseudo.alt)*d$samples$norm.factors", the pseudocounts Sum of each > library are very similar to each other as expected. > To follow this point I have another question about the meaning of the above > formula. What does it mean in mathematics or/and biology, if any? > If I call the pseudo-counts (from d$pseudo.alt) as the > library-size-adjusted-counts (right ?), can I call the new number > "d$pseudo.alt*d$samples$norm.factors" as normalized-counts? The calculation > of pseudocounts is explained in the User's Guide, but not this one, except > one sentence that might be related in the man page of calcNormFactors(): "For > symmetry, normalization factors are adjusted to multiply to 1". I do not > feel very clear with it. > > The reason I want to get this "normalized number" is to compare the relative > expression level of each gene across the conditions. I am somehow confused > with the difference between the library size adjustment and the normalization > (This is mentioned on page 3 of the user's guide), when I was trying to > convince my colleague the pseudo-counts of a gene should be used to compare > the relative expression level, but this formula integrates both library size > and the normalization factor, which seems somehow redundant, am I right? > > Another general question is the row names of the data. I found after the > DGEList object was created, the row names are gone. Before the analysis I > assigned the row names to the IDs, but the results are labeled with tag.1, > tag.2, tag.53 etc when the analysis is done: > > head(de.com.ZO$table) > logConc logFC p.value > tag.1 -17.48306247 2.1215798358 2.723289267e-05 > tag.2 -17.06332866 0.6656604697 1.705717572e-01 > tag.3 -11.99482641 -0.3162543327 4.915230723e-01 > tag.4 -15.76854867 -0.5882280375 2.121440211e-01 > tag.5 -15.22523652 1.9007994967 7.434382812e-05 > > How to o keep the rownames always in position? This may be related to R data > manupilation, but with edgeR I felt lost, especially after I filtered out the > extreme low cpm genes as the dataset is different from the original one. > > ################This is what I used as copied from the guide############### > countsTable <- read.delim("Deep_seq_data.csv", header=T, stringsAsFactors=T) > > # Remove the first column which is FeatureID as AGI number > d<-countsTable[, -1] > > conditions<-rep(c("Zygote", "Octant", "Globular", "Heart", "Torpedo", "Bent", > "Mature"), each=2) > > #set the row name for database, believed to be mySQL > rownames(d)<-countsTable[,1] > > # make a DGE object > d<-DGEList(counts=countsTable[,-1], group=conditions) > > #check the counts > dim(d) > head(d$counts) > head(d$samples) > > # Filter genes with >=1 counts per million, in at leats 2 samples > d<-d[rowSums(1e+06*d$counts/expandAsMatrix(d$samples$lib.size, > dim(d))>=1)>=2,] > > #get the gene number satisfying the above condition and for display use of > topTag > dim(d) > d > > d > An object of class "DGEList" > $samples > group lib.size norm.factors > Zygote1 Zygote 21012147 1 > Zygote2 Zygote 19924212 1 > Octant1 Octant 26002900 1 > Octant2 Octant 9660245 1 > Globular1 Globular 17139388 1 > 9 more rows ... > > $counts > Zygote1 Zygote2 Octant1 Octant2 Globular1 Globular2 Heart1 Heart2 > Torpedo1 Torpedo2 Bent1 Bent2 Mature1 Mature2 > [1,] 40 42 322 158 616 402 1511 1897 > 1122 640 1121 351 0 36 > [2,] 115 68 270 123 94 37 403 517 > 83 24 166 66 2 22 > [3,] 4228 4340 4794 3678 1527 1146 960 1491 > 747 465 2831 2138 11880 24451 > [4,] 242 441 369 223 72 40 628 374 > 223 111 1796 426 106 133 > [5,] 220 204 1427 699 1233 332 1108 1011 > 474 284 499 129 96 196 > 20587 more rows ... > > $all.zeros > [1] FALSE FALSE FALSE FALSE FALSE > 20587 more elements ... > ...... > all other steps went well without any problem except the following one > ...... > detags.com.ZO <- rownames(topTags(de.com.ZO, n=20592)$table) #n=20592 is > the total row number in the analysis > > #Always got problem with this line even I declare the total number of rows > in full as above. > > d$counts[detags.com.ZO, ] > Error: subscript out of bounds > > sessionInfo() > R version 2.13.0 (2011-04-13) > Platform: x86_64-pc-linux-gnu (64-bit) > locale: > [1] LC_CTYPE=en_CA.UTF-8 LC_NUMERIC=C LC_TIME=en_CA.UTF-8 > LC_COLLATE=en_CA.UTF-8 > [5] LC_MONETARY=C LC_MESSAGES=en_CA.UTF-8 > LC_PAPER=en_CA.UTF-8 LC_NAME=C > [9] LC_ADDRESS=C LC_TELEPHONE=C > LC_MEASUREMENT=en_CA.UTF-8 LC_IDENTIFICATION=C > attached base packages: > [1] stats graphics grDevices utils datasets methods base > other attached packages: > [1] ALL_1.4.7 Biobase_2.12.1 limma_3.8.2 edgeR_2.2.5 > loaded via a namespace (and not attached): > [1] tools_2.13.0 > > ############################################################## > > My questions are long, but I feel after these two questions get clear, I am > ready with edgeR. > > Thank you! > > Yifang > > Yifang Tan > > > > Subject: Re: [Bioc-sig-seq] edgeR common library size question. > From: [email protected] > Date: Wed, 22 Jun 2011 12:28:56 +1000 > CC: [email protected] > To: [email protected] > > Hi Yifang > > I have discussed this issue with Gordon and Mark R---briefly, this is not > something to worry about. We do not expect the library sizes of the > pseudocounts necessarily to be approximately equal when TMM normalization (or > another type of normalization factor) is applied. That comment in the User's > Guide dates back to the pre-TMM normalization days and is no longer relevant. > I will update the User's Guide for the next release to remove that confusing > statement. Sorry for the angst that note in the Guide has caused. > > In answer to your specific questions: > 1) Best way to check that the normalization has performed well is to look at > MA-plots (plotSmear in edgeR). > > 2) It does not matter that the normalized library size of the conditions > differs from the common.lib.size. > > 3) Yes, the results are still meaningful even if the library sizes of the > pseudocounts are different. > > 4) I suspect that the differences in library sizes of the pseudocounts are > caused by the same thing that causes differences in normalization factors: > sets of genes that are highly expressed in a given sample but not in others. > For some reassurance, you could look at: > colSums(d$pseudo.alt)*d$samples$norm.factors > These values should be quite similar (do not worry too much if they differ), > which shows the effect of the normalization factors on the effective library > size. > > 5) We routinely remove genes with very low expression levels in our analysis > and I would recommend it generally. In your experiment you have a minimum > group size of two, so I would recommend keeping genes that are expressed at a > level of 0.5 counts per million or greater in at least two samples. This is > almost exactly what you did in your earlier email. Removing genes with <=40 > counts in >= 6 samples is probably overdoing it. > > Example code (in the next release there will be a convenience function to > calculate counts per million easily from a DGEList object): > cpm <- 1e06*t( t(d$counts) / d$samples$lib.size ) > d <- d[ rowSums( cpm > 0.5 ) >= 2, ] > > Mark provided me with a really nice reproducible example (thanks Mark!) that > illustrates how and why TMM normalization works and also why we do not expect > the library sizes of the pseudocounts to be equal after normalization factors > are applied. Try running the code below for yourself: > > ======================== > library(edgeR) > > f <- url("http://bioinf.wehi.edu.au/folders/tmm_rnaseq/LK_data.RData";) > load(f) > close(f) > > D <- as.matrix(MA.subsetA$M) > g <- as.character(MA.subsetA$genes$EnsemblGeneID) > > source("http://bioinf.wehi.edu.au/folders/tmm_rnaseq/functions.R";) > > libSizes <- colSums(D) > foldDiff <- 3 > pUp <- .8 > pDifferential=0.1 > > xx <- generateDataset(commonTags=2e4, uniqueTags=c(3000,100), > foldDifference=foldDiff, > pUp=pUp, pDifferential=pDifferential, > empiricalDist=D[,1], libLimits=rep(1e6,2)) > > d <- DGEList(counts=xx$DATA, group=c(1,2)) > d1 <- calcNormFactors(d) > > par(mfrow=c(1,2)) > plotSmear(d, ylim=c(-5,5)) > plotSmear(d1, ylim=c(-5,5)) > > d <- estimateCommonDisp(d) > colSums(d$pseudo.alt) > d1 <- estimateCommonDisp(d1) > colSums(d1$pseudo.alt) > ========================== > > Hope that clarifies a few things for you. > > Best wishes > Davis > > P.S. In general it is good form not to re-post the same question again on the > list. We receive and read questions on the BioC mailing lists, but sometimes > it can be a few days before we can make time to respond. > > > > On Jun 22, 2011, at 7:50 AM, yifang tan wrote: > > > Hi Davis/Gordon: > I posted my question here again hope you can see it. > When I tried edgeR and met a problem with the number of pseudocounts > for each library after normalization, which should come to close numbers. > This have been addressed in edgeR > several times that "the total counts in each libray of the pseudocounts > agrees well with the common library size" (page 27 & 44 of the > user's guide), but my result are quite different between treatments although > for the replicates within treatment the pseudocounts are very similar. I > can't get to the common.lib.size for each treatment after I tried several > methods (TMM, RLE and quantile). > 1) Did I miss anything during my run with edgeR? How can I assure the > normalization went well? > > 2) Does the normalized library size of the conditions matter or NOT, if they > are different from the common.lib.size? > > > 3) Is the result still meaningful even the library sizes of pseudocounts are > different? > > > 4) What could probably be the reason(s) to cause the library sizes of > pseudocounts so different? > > > 5) Should I remove the smaller number reads as some other people do? > After I removed the smaller numbers of counts (<=40 in >=6 out of > 14 samples), the normalized library sizes become very similar. > > > I can feel my lack of mathematics for the packages. I attach part of my code > here. > > --------------------------------------------------------------------- > d$samples$lib.size > #"Zygote1", 21012147 > "Zygote2", 19924212 > "Octant1", 9660245 > "Octant2", 26002900 > "Globular1",17139388 > "Globular2", 7649319 > "Heart1", 16430105 > "Heart2", 20101956 > "Torpedo1", 12920266 > "Torpedo2", 6306742 > "Bent1", 44241095 > "Bent2", 20094409 > "Mature1", 15166090 > "Mature2", 23203758 > > d$common.lib.size > [1] 16554344.47 > > colSums(d$pseudo.alt) > # Zygote1 21523774.62 > Zygote2 21638415.63 > Octant1 14533481.82 > Octant2 12046955.46 > Globular1 18920316.62 > Globular2 18439528.30 > Heart1 11754608.30 > Heart2 12759230.11 > Torpedo1 11248245.52 > Torpedo2 11410667.92 > Bent1 16101723.65 > Bent2 17980670.24 > Mature1 26785396.02 > Mature2 27067289.80 > # > > sessionInfo() > R version 2.13.0 (2011-04-13) > Platform: x86_64-pc-linux-gnu (64-bit) > locale: > [1] LC_CTYPE=en_CA.UTF-8 LC_NUMERIC=C LC_TIME=en_CA.UTF-8 > LC_COLLATE=en_CA.UTF-8 > [5] LC_MONETARY=C LC_MESSAGES=en_CA.UTF-8 > LC_PAPER=en_CA.UTF-8 LC_NAME=C > [9] LC_ADDRESS=C LC_TELEPHONE=C > LC_MEASUREMENT=en_CA.UTF-8 LC_IDENTIFICATION=C > attached base packages: > [1] stats graphics grDevices utils datasets methods base > other attached packages: > [1] ALL_1.4.7 Biobase_2.12.1 limma_3.8.2 edgeR_2.2.5 > loaded via a namespace (and not attached): > [1] tools_2.13.0 > --------------------------------------------------------------------- > [[elided Hotmail spam]] > > > Yifang > > > Yifang Tan > > [[alternative HTML version deleted]] > > _______________________________________________ > Bioc-sig-sequencing mailing list > [email protected] > https://stat.ethz.ch/mailman/listinfo/bioc-sig-sequencing > > ------------------------------------------------------------------------ > Davis J McCarthy > Research Technician > Bioinformatics Division > Walter and Eliza Hall Institute of Medical Research > 1G Royal Parade, Parkville, Vic 3052, Australia > [email protected] > http://www.wehi.edu.au > > > > > ______________________________________________________________________ > The information in this email is confidential and intended solely for the > addressee. > You must not disclose, forward, print or use it without the permission of the > sender. > ______________________________________________________________________ > > ------------------------------------------------------------------------ > Davis J McCarthy > Research Technician > Bioinformatics Division > Walter and Eliza Hall Institute of Medical Research > 1G Royal Parade, Parkville, Vic 3052, Australia > [email protected] > http://www.wehi.edu.au > > > > > ______________________________________________________________________ > The information in this email is confidential and intended solely for the > addressee. > You must not disclose, forward, print or use it without the permission of the > sender. > ______________________________________________________________________ ------------------------------------------------------------------------ Davis J McCarthy Research Technician Bioinformatics Division Walter and Eliza Hall Institute of Medical Research 1G Royal Parade, Parkville, Vic 3052, Australia [email protected] http://www.wehi.edu.au ______________________________________________________________________ The information in this email is confidential and intended solely for the addressee. You must not disclose, forward, print or use it without the permission of the sender. ______________________________________________________________________ [[alternative HTML version deleted]] _______________________________________________ Bioc-sig-sequencing mailing list [email protected] https://stat.ethz.ch/mailman/listinfo/bioc-sig-sequencing
