DESeq: question about baseMean. Also, replicates.

[Azazel]

One question about replicates in DESeq.

I read "Analysing RNA-Seq with the DESeq package", version from April 21, 2011.

In the "Quick Start" section of abovementioned document, they strongly advice to sum-up "purely technical" replicates and have only biological replicates in separate columns. However, in my experiment it's not clear to me if some of the replicates are "purely technical" or rather biological: I am analyzing different brain tissue samples in treated and untreated states. I have samples where the RNA-preparation has been done in duplicate from the same biological tissue, then sequenced independently. Is this to be considered technical or biological in the context of DESeq?

[natstreet]

2.) In the "Quick Start" section of abovementioned document, they strongly advice to sum-up "purely technical" replicates and have only biological replicates in separate columns. However, in my experiment it's not clear to me if some of the replicates are "purely technical" or rather biological: I am analyzing different brain tissue samples in treated and untreated states. I have samples where the RNA-preparation has been done in duplicate from the same biological tissue, then sequenced independently. Is this to be considered technical or biological in the context of DESeq?

The answer to this shouldn't have anything to do with DESeq - it's the same issue regardless of the software used. I would say the answer lies in part of your question

in duplicate from the same biological tissue

To me it is then clear - this is a technical replicate and it would be wrong to classify those duplicate sample preps as independent biological replicate as they are from the same individual.

[ning]

I think maybe you could roughly look at the mean and var of each gene. If the var/mean is much higher than 1, then you may treat them as "bilogical rep"? And if the var/mean is around 1, you may treat them as "tech rep".. Since people saw technical reps are poisson distributed. And for biological reps, the var of the poisson is overdispersed.

[natstreet]

I think maybe you could roughly look at the mean and var of each gene. If the var/mean is much higher than 1, then you may treat them as "bilogical rep"? And if the var/mean is around 1, you may treat them as "tech rep".. Since people saw technical reps are poisson distributed. And for biological reps, the var of the poisson is overdispersed.

I'm not sure what the relationship between var and mean has to do with whether the samples are classified as biological or technical replicates. A biological replicate requires that the samples are independent. Extracting a second sample from the same individual will allow you to determine technical bias but it does not increase knowledge about how biologically reproducible a result it. Simon Andrews has posted some good answers in relation to this in other threads.

[ning]

I'm not sure what the relationship between var and mean has to do with whether the samples are classified as biological or technical replicates. A biological replicate requires that the samples are independent. Extracting a second sample from the same individual will allow you to determine technical bias but it does not increase knowledge about how biologically reproducible a result it. Simon Andrews has posted some good answers in relation to this in other threads.

Thanks for the comment!
I'm just kind of thinking when we run DESeq or edgeR, both of them are assuming that different lanes follow a Negative Binomial distribution with the same "overdisperse parameter" corss all the lanes.
If there are 4 lanes Lane1... Lane4. If Lane 1 and Lane2 are pure tech reps, then the "overdisperse parameter" of Lane1 and Lane2 will be small. But If we look at Lane1 vs Lane3 vs Lane4, the "overdisperse parameter" will be much larger. Then the "same overdisperse parameter" assumption won't hold if we analyze 4 lanes together. and the estimator of the overdisperse parameter will be pretty poor. That's why we need to sum-up "pure tech reps" in my understanding.
So I'm thinking if we could observe the var of Lane1 and Lane2 is the same as the var of Lane1, Lane3, Lane4, I'll treat Lane1 and Lane2 as "Not-pure" technical reps since the "same overdisperse parameter" assumption holds. So I won't sum them up.
Although maybe it's hard to get a good estimator of var if we only have two lanes...
Does this make sense?

Thanks

[Simon Anders]

A while ago, I've given a more comprehensive explanation of the rationale behind the treatment of technical replicates on the Bioconductor mailing lits. Maybe it is of interest in the current context:

https://stat.ethz.ch/pipermail/bioconductor/2010-November/036395.html