Another DESeq question

[shurjo]

Hi Simon,

First of all, many thanks for DESeq. I have found it wonderfully easy to work with, even with my beginner's knowledge of R. My understanding of statistics is limited to a couple of graduate-level courses, so I would really appreciate your input on the following questions, which arise from using DESeq on my data. My project comprises 14 individuals (7 case-control pairs) and we have ~40 million Illumina transcriptome reads from each of these people. I used BedTools to overlap these reads with the UCSC knownGene annotations, and ran DESeq on the count data.

Question 1:
The ECDF curves lie above the diagonal in both cases, so I used a correction factor of 0.7, which brings them closer. I have attached both the uncorrected and corrected curves below, as well as the plots of the fit of the variance functions. Can you tell me if, in your opinion, the correction factor seems adequate?

Question 2:
On page 10 of the manual, you mention that any hit with a very large value in any one of the last two columns in the differential expression output file should be checked carefully to exclude false hits due to “variance outliers”. Do you have a ballpark figure in mind when you say "very large value"? In my data, the values range from 0.003 - 214, so I am confused what to use as a cutoff.

Any advice you have to offer would be very welcome.

Thanks,

Shurjo

[Simon Anders]

Hi Shurio

Question 1:
The ECDF curves lie above the diagonal in both cases, so I used a correction factor of 0.7, which brings them closer. I have attached both the uncorrected and corrected curves below, as well as the plots of the fit of the variance functions. Can you tell me if, in your opinion, the correction factor seems adequate?

I'm afraid I wrote quite some nonsense last week, when I amended the package vignette last week to describe the variance adjustment factors, see my post here.

Now correcting my mistake of last week, i reverse my advice: If the curves are above the green line, use the fit as it is. Bringing them down onto the green line may DESeq become anti-conservative due to the curves now being partly too low.

Question 2:
On page 10 of the manual, you mention that any hit with a very large value in any one of the last two columns in the differential expression output file should be checked carefully to exclude false hits due to “variance outliers”. Do you have a ballpark figure in mind when you say "very large value"? In my data, the values range from 0.003 - 214, so I am confused what to use as a cutoff.

I am still trying to find a good general answer to this question. I have an idea, and I may add a feature quite soon for this. Until then, you may want to plot a histogram of variance values, e.g., with

hist( res$resVarA, breaks=100 )
hist( res$resVarB, breaks=100 )

You will probably notice that 99% or so of the values are below 10 or 20 or so and only very few as far up at 214. This histogram will help you judge which values are clearly outliers, i.e., out of the range of most of the values.

For two or three replicates, a cut-off around 15 or 20 seems to be good, but as you have many more replicates, your histogram may look different.

Cheers
Simon

[shurjo]

Hi Simon,

Your advice makes perfect sense, thanks.

Regards,

Shurjo