DESeq without biological replicates (dataset: Marioni et al.)

[Andrea Apolloni]

Hello,
I'm trying to use DESeq (version 1.6.1) for differential analysis of the dataset of Marioni et al.
This dataset contains the counts of 32000 genes sequenced from liver and kidney cells of only one patient. For each condition (liver and kidney), the authors sequenced the samples on different lanes of the machine. This results in replication techniques, but not biological replicates.

Following the directions of the vignette, I summed the counts of technical replicates of liver and kidney. So, now I have a sample for kidney and a sample for liver, for each gene.

I'm trying to use the estimateDispersions function with the parameters method="blind" and sharingMode="fit-only".
By doing so I get a warning: "In parametricDispersionFit(means, disps) : Dispersion fit did not converge."
If I try to use the estimateDispersions function with the parameters method="blind", sharingMode="fit-only" and also fitType="local", the warning disappears.

Is this a correct choice of parameters, with this dataset?

Thanks

[Simon Anders]

Yes, even though the results will still not be very meaningful.

[Andrea Apolloni]

Thanks for the clarification, Simon.

So fitType="local" may be safely used when fitType="parametric" does not fails?

My doubt is because in the vignette (chapter "Working without Any replicates") you don't sum the counts of the pasilla dataset, but you arbitrarily choose to use only two samples (one per condition).
You make this choice to "simulate" a dataset in which there isn't any kind of replicates?

[Simon Anders]

So fitType="local" may be safely used when fitType="parametric" does not fails?

Yes. This is what the warning recommends, and, actually, I am not even sure yet which fit type I'd prefer.

My doubt is because in the vignette (chapter "Working without Any replicates") you don't sum the counts of the pasilla dataset, but you arbitrarily choose to use only two samples (one per condition).
You make this choice to "simulate" a dataset in which there isn't any kind of replicates?

Exactly.