SEQanswers

Go Back   SEQanswers > Bioinformatics > Bioinformatics



Similar Threads
Thread Thread Starter Forum Replies Last Post
EdgeR and Deseq Normalization problems moriah Bioinformatics 4 03-26-2014 03:32 AM
sRNA RNAseq Normalization Question rndouglas Bioinformatics 0 10-29-2012 03:22 PM
EdgeR Normalization Question R_user RNA Sequencing 1 05-07-2012 10:14 PM
edgeR Normalization question polsum Bioinformatics 1 05-07-2012 12:21 AM
NORMALIZATION Illumina DESeq HTFRIEDELFF Bioinformatics 1 12-15-2010 12:47 AM

Reply
 
Thread Tools
Old 05-13-2013, 10:31 AM   #1
turnersd
Senior Member
 
Location: Charlottesville, VA

Join Date: May 2011
Posts: 112
Default DESeq Normalization Question

I'm sure this issue has come up before, but I couldn't find an appropriate thread or answer either here or on the Bioconductor mailing list.

What feature of the data or the distribution of counts among my samples can cause the sizeFactors to vary much more than the raw counts / library sizes?

More detail: I'm using DESeq to analyze RNA-seq data mapped with STAR, counted with htseq-count. Comparing the "doubleTerm" samples to the "wt" samples, there are many genes that appear downregulated. While these samples were sequenced, on average, to a similar sequencing depth, the normalization factors are much smaller for WT, resulting in much larger normalized counts, resulting in more apparently downregulated genes in doubleTerm vs WT.
Code:
> cds <- newCountDataSetFromHTSeqCount(sampleTable=sampleTable, directory=directory)
> cds <- estimateSizeFactors(cds)
> cds <- estimateDispersions(cds)
> data.frame(sizefactors=sizeFactors(cds), rawcounts=colSums(counts(cds, normalized=FALSE)))
                sizefactors rawcounts
S01_wt1           0.9016089  23466349
S02_wt2           0.7679168  22428603
S03_wt3           0.7952564  19841959
S04_wt4           0.7839629  18363384
S05_pten8w1       1.0301769  20859853
S06_pten8w2       0.9949514  16809588
S07_pten8w3       0.9425865  16731071
S08_pten22w1      1.0826846  18906329
S09_pten22w2      1.1640354  20164026
S10_pten22w3      1.0111748  17306468
S11_double8w1     0.7949001  17671986
S12_double8w2     1.4509978  23673557
S13_double8w3     1.1703853  22127841
S14_doubleterm2   1.0786455  19063010
S15_doubleterm4   1.1265935  19279814
S16_doubleterm6   1.3059472  22750403
Thanks.

Stephen

Code:
> sessionInfo()
R version 3.0.0 (2013-04-03)
Platform: x86_64-apple-darwin10.8.0 (64-bit)

locale:
[1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8

attached base packages:
[1] parallel  stats     graphics  grDevices utils     datasets
methods   base

other attached packages:
[1] DESeq_1.12.0         lattice_0.20-15      locfit_1.5-9
Biobase_2.20.0
[5] BiocGenerics_0.6.0   edgeR_3.2.3          limma_3.16.2
BiocInstaller_1.10.1

loaded via a namespace (and not attached):
 [1] annotate_1.38.0      AnnotationDbi_1.22.3 DBI_0.2-6
DESeq2_1.0.9
 [5] genefilter_1.42.0    geneplotter_1.38.0   GenomicRanges_1.12.2
grid_3.0.0
 [9] IRanges_1.18.0       RColorBrewer_1.0-5   RSQLite_0.11.3
splines_3.0.0
[13] stats4_3.0.0         survival_2.37-4      tools_3.0.0
XML_3.95-0.2
[17] xtable_1.7-1
turnersd is offline   Reply With Quote
Old 05-13-2013, 12:08 PM   #2
kopi-o
Senior Member
 
Location: Stockholm, Sweden

Join Date: Feb 2008
Posts: 319
Default

One thing that could cause this is extreme overrepresentation of one or a handful of genes, like hemoglobin in whole-blood samples for instance. Or maybe rubisco in plants. Partial/failed rRNA removal? I would look for something that seems to dominate the RNA pool.
kopi-o is offline   Reply With Quote
Old 05-20-2013, 01:32 AM   #3
Wolfgang Huber
Senior Member
 
Location: Heidelberg, Germany

Join Date: Aug 2009
Posts: 108
Default

Dear Stephen,

how does the 'pairs' (or LSD::heatpairs) plot look like, or the pairwise MA plots? You could also try the arrayQualityMetrics report. From this, problems such as suggested by kopi-o might become apparent.

Best wishes
Wolfgang.
__________________
Wolfgang Huber
EMBL
Wolfgang Huber is offline   Reply With Quote
Old 11-24-2014, 10:57 PM   #4
bpb9
Member
 
Location: NYC

Join Date: Aug 2012
Posts: 24
Default

I have what is probably a naive question about DESeq normalization. The manual says that by dividing the count by the size factor, one makes samples comparable. Does this include comparable for, say, eQTL analysis? I randomly pulled some genes from counts(cds,normalized=TRUE) and plotted their distributions, and they are not all normalliy distributed. Some look normal-ish (bottom row), but I still don't know if I'd consider them normal; others are very clearly following other distributions (top row). If I want to identify eQTLs from my RNAseq data (as I also have genotype data on those individuals), then all the gene read counts need to be transformed to follow a normal distribution before I can test for eQTLs.

So my question is, is the command counts(cds,normalized=TRUE) designed to transform the raw reads counts into expression levels that follow a normal distribution? If so, why do some of my genes not look normally distributed? If not, how could I transform my countDataSet so that each gene follows a normal distribution? If my dataset consists of expression data from three time points, could that be messing up what would otherwise be a normally distributed gene? Figure of some distributions for genes from counts(cds,normalized=TRUE) are below. Thanks for any suggestions!
bpb9 is offline   Reply With Quote
Old 11-25-2014, 10:06 PM   #5
gringer
David Eccles (gringer)
 
Location: Wellington, New Zealand

Join Date: May 2011
Posts: 817
Default

I would agree with kopi-o. The normalisation isn't based on the total counts, it's based on the distribution of counts across all genes. If you want a slightly more accurate comparison to size factors (but fairly easy to calculate), try splitting the count data into quantiles, or look at (say) the 75th percentile count instead of the total count.
gringer is offline   Reply With Quote
Old 11-25-2014, 10:19 PM   #6
bpb9
Member
 
Location: NYC

Join Date: Aug 2012
Posts: 24
Default

I came across this lovely thing: http://www.bios.unc.edu/research/gen..._eQTL/faq.html

"Outliers in expression data are usually harder to deal with. The accepted remedy by the GTEx consortium is the transformation of the measurements for each gene into normally distributed while preserving relative rankings. The target distribution may be the standard normal distribution or the normal distribution the mean and spread of the original measurements. Here is the code for such transformation:"

for( sl in 1:length(gene) ) {
mat = gene[[sl]];
mat = t(apply(mat, 1, rank, ties.method = "average"));
mat = qnorm(mat / (ncol(gene)+1));
gene[[sl]] = mat;
}
rm(sl, mat);

I used my normalized DESeq count data as input, then used the program to transform each gene to a normal distribution of expression. Comparing before and after transformation for a few genes, they certainly look normal.



The program claims to have been used successfully to identify eQTLs in RNAseq data. Whether or not my using this approach for eQTLs turns out to be biologically relevant, informative, or correct is yet to be determined. Has anyone tried other transformations of RNAseq for eQTL analysis?
bpb9 is offline   Reply With Quote
Reply

Tags
bioconductor, deseq, normalization, rna-seq, rstats

Thread Tools

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off




All times are GMT -8. The time now is 02:00 PM.


Powered by vBulletin® Version 3.8.9
Copyright ©2000 - 2018, vBulletin Solutions, Inc.
Single Sign On provided by vBSSO