[Simon Anders]

If you have m samples, then the maximum possible coefficient of variance is achieved if a gene appears in only one samples and has zero counts in the (m-1) other samples. In these cases, you will see a dispersion value close to m. So, all these points in the top left of your plots are likely genes that appear in only one or very few samples and are absent in most other samples. So, if you look at a scatter plot of one sample versus another, most of these genes will just vanish in the bottom-left zero-zero corner. This is why you didnt see them there.

I don't know what "cdsFilt" is, but removing these genes did make the plot more normal looking. However, note that if you zoomed in your previous plot into the region with mean > 10, it looks quite the same as the new plot. So your results shouldn't change much, i.e., the filtering might not even have been necessary.

[bpb9]

Hi Simon, by cdsFilt I mean I did the following:

cdsFilt:

rs <- rowSums ( counts ( cds ))
use <- (rs > quantile(rs, 0.4))
table(use)
FALSE TRUE
10,000 15,243
#Proceed with only filtered data (15,243 genes)
cdsFilt <- cds[ use, ]

cdsFilt <- estimateDispersions( cdsFilt )

#View dispersion estimates
str( fitInfo(cdsFilt) )

#Plot filtered dispersions
plot(
rowMeans( counts( cdsFilt, normalized=TRUE ) ),
fitInfo(cdsFilt)$perGeneDispEsts,
pch = '.', log="xy"
)
xg <- 10^seq( -.10, 10, length.out=300 )
lines( xg, fitInfo(cds)$dispFun( xg ), col="red" )

If I am going to use this data to test for differential expression and eQTLs, then I think removing the low expression transcripts, while not necessary, would be helpful, as it reduces multiple test correction.

[smurmu]

Can anyone please help in normalizing rna seq data using EDASeq??
Thanks