edgeR dispersion value (common BCV (square- root-dispersion)) when no replicates in dataset

Question

[fatimarasool135]

I am performing several analysis of differential expression (using edgeR). Reading edgeR manual I am not sure about BCV value that I should use. I am comparing libraries of the same species (non model species) and i have no biological replicates.

In the manual you say that "Typical values for the common BCV (square- root-dispersion) for datasets arising from well-controlled experiments are 0.4 for human data, 0.1 for data on genetically identical model organisms or 0.01 for technical replicates."

The species that I am studying is wheat . Can you please help me so I can decide which dispersion value to use.

[ashish]

I have done the same thing on wheat once. I tried multiple approaches but without replicates it is not very useful. The best results I found which matched wet lab data were from GFOLD. Here also you will not get any statistics, just ranked genes. You can use the top and below 200/300/400 genes. This will just give you an idea about major changes but they must be verified by other downstream experiments.

[ATpoint]

Without replicates it is completely arbitrary what you use. The true value could be much smaller or notably inflated due to experimental conditions. Unreplicated experiments are not suited for differential analysis. What you can do is rank fold changes after transforming data with e.g. vst or rlog from DESeq2 but don't make up any dispersion values, as there is absolutely no data-driven evidence that the value is correct. You would be lying to yourself. The damage has been done when designing unreplicated experiments, not much you can do about it.

[fatimarasool135]

Hi ATpoint, I have read vst in the manual of DEseq2. but i could not understand itkindly write command I have compared the samples to each other by using the Deseq2 tool and have list of genes with parameters generated by deseq2 tool. Now i have no chance to take biological replicates. You have told me that I have to rank the fold changes. then how i can i do it?

If I rank it with fold change then it will be differential analysis ?

[ATpoint]

You can normalize your counts with vst or rlog. These commands take a DESeq2 object or a count matrix as input. The DESeq2 manual covers how to do that. After that you divide the normalized counts , e.g. counts1/counts2 to get fold changes. These you can rank from positive to negative and then e.g. take those as candidates which have a FC of e.g. > 2. This is not differential analysis, as you do not have replicates. You can take these candidates for downstream experiments but be aware, it is candidates and no analysis in the world will give solid results from unreplicated data.