SEQanswers

SEQanswers (http://seqanswers.com/forums/index.php)
-   RNA Sequencing (http://seqanswers.com/forums/forumdisplay.php?f=26)
-   -   Questionable diagnostic plots from Cummerbund/DESeq2 (http://seqanswers.com/forums/showthread.php?t=50977)

DPCook 03-12-2015 08:01 AM

Questionable diagnostic plots from Cummerbund/DESeq2
 
5 Attachment(s)
Hi everyone,

I'm analyzing some RNA-seq data for a colleague and some of the results/diagnostic plots have raised some caution flags in my mind. I'll include the code I used to avoid going back and forth. Cliffnotes: Getting a large number of differentially expressed genes (which may or may not be normal?), large log2FCs, and diagnostic plots don't quite look typical.

Experimental Design:
RNA-seq before and after inducing differentiation(GNP -> GC). Two biological replicates for each condition.

Alignment & Counts (done for each of the four samples):
Code:

tophat -p 8 -G Mus.musculus.NCBIM37.65.gtf -o tophat_dir --no-novel-juncs mm9 GC2_R1.fastq GC2_R2.fastq
samtools sort -n accepted_hits.bam sorted.GC2
htseq-count -f bam sorted.GC2 Mus.musculus.NCBIM37.65.norRNA.gtf > GC2.counts.txt ##annotation lacks rRNA just in case

I then used DESeqDataSetFromHTSeqCount to get the summary and files set up:
Code:

sampleName        fileName condition
1  GC2.counts.txt  GC2.counts.txt        GC
2  GC3.counts.txt  GC3.counts.txt        GC
3 GNP2.counts.txt GNP2.counts.txt      GNP
4 GNP3.counts.txt GNP3.counts.txt      GNP

> dds
class: DESeqDataSet
dim: 37651 4
exptData(0):
assays(3): counts mu cooks
rownames(37651): ENSMUSG00000000001 ENSMUSG00000000003 ... ENSMUSG00000093788
  ENSMUSG00000093789
rowData metadata column names(27): baseMean baseVar ... deviance maxCooks
colnames(4): GC2.counts.txt GC3.counts.txt GNP2.counts.txt GNP3.counts.txt
colData names(2): condition sizeFactor

Ran DESeq and checked results:
Code:

dds <- DESeq(dds)
res <- results(dds)

> summary(res)

out of 24178 with nonzero total read count
adjusted p-value < 0.1
LFC > 0 (up)    : 5752, 24%  ##~1000 LFC > 2
LFC < 0 (down)  : 5786, 24% ##~1000 LFC < 2
outliers [1]    : 0, 0%
low counts [2]  : 4687, 19%
(mean count < 1.7)
[1] see 'cooksCutoff' argument of ?results
[2] see 'independentFiltering' argument of ?results

##MA plot
plotMA(res, main="DESeq2", ylim=c(-10,10)) ##Large number of significant genes with high mean exp value

##P value dist.
hist(res$pval, breaks=100) ##Very few genes with high p-value (looks similar with padj)

##Dispersion plot
plotDispEsts(dds) ##Not really sure if there's anything strange here

##Per gene standard deviation plots
#Script exactly as presented in DESeq2 vignette: shifted logarithm log2(n + 1) (left), the regularized log transformation(center), and the variance stabilizing transformation (right)
#Some really high SDs

##Euclidian distances
distsRL <- dist(t(assay(rld)))

What's reassuring is that the differentially expressed genes are enriched for relevant GO terms, but the plots are just looking different from others I have seen. Could anyone shed some light on whether there is anything to be alarmed about?

EDIT: Just realized I mentioned cuffdiff/cummeRbund in the title but didn't include anything from it. Volcano plot from there showed results consistent with the MA plot here.

David Cook
MSc. Candidate, University of OTtawa

Michael Love 03-12-2015 06:11 PM

PCA plot is also nice for seeing sample distances. From the distplot these conditions look very distinct. Why don't you consider testing at a higher threshold than |LFC| > 0, as this seems to be achieved by many genes? See the lfcThreshold argument of ?results. The reasoning is described in the "Specifying minimum effect size" section of the paper.

DPCook 03-13-2015 04:49 AM

1 Attachment(s)
Thanks for the reply Michael. I adjusted the lfc threshold and it certainly made the list a bit more manageable (LFC > 1.5, FDR=0.05 yielded about 800 DEGs in both directions). I guess I just wasn't expecting such large differences between the two conditions and was concerned that I was missing some obvious artefact that could cause it.

I also ran the PCA following regularize log transformation to look at distances. PC1 is apparently capturing 100% variance and splits the two conditions. I'm no expert, so correct me if I'm wrong, but I suppose this supports that idea that the results are just the product of very distinct conditions because the variability between biological replicates is negligible (at least relative to the differences between conditions).

Thanks!

netgear extender setup 11-06-2019 01:15 AM

Thanks for sharing. We provide full support for all your Arlo devices, including guidance for your netgear extender setup . So if you are having issues with connecting to the Wifi or configure settings on the Arlo app, then contact us using our live chat services or our email. You can also call us using our customer support phone number.


All times are GMT -8. The time now is 05:08 PM.

Powered by vBulletin® Version 3.8.9
Copyright ©2000 - 2020, vBulletin Solutions, Inc.