Help required with edgeR contrasts

[Bouteille]

Hi,

I am using edgeR for analyzing a batch of RNA-Seq data made of differement cell compartment extracts (nuclear and cytoplasmic) before stimulation (00h) and several timepoints after stimulation (02h and 04h). Each condition has two replicates.
The goal is to analyze whether the stimulation has an effect on the nuclear export of RNAs.

For that, I would like to study :
-Which annotations are significantly compartimented in cytoplasm or nucleus for each timepoint (including 00h)
-The effect of stimulation (using 00h as reference) on the compartimentalization
....first by comparing 00h to individual timepoints 02h and 04h separately
....then trying to see what are the effect of stimulation independently of the time after stimulation by comparing 00h to the mix of all timepoints after stimulation

I built a matrix containing counts of all experiments and a design matrix 'edgerGlmDesign' as follow (no intercept) :

>colnames(edgerGlmDesign)
"Cyto.00h" "Cyto.02h" "Cyto.04h" "Nucl.00h" "Nucl.02h" "Nucl.04h"

Before using the results of the analysis I would like to be sure that I am indeed comparing the right things (I doubt it actually) with the following setup :

   
    # Prepare the 'contrasts' summarizing the conditions we want to compare
    myContrasts <- makeContrasts(
            # First Nuclear and Cytoplasmic will be compared for each timepoint individually
            NuclearVsCytoplasmic_00h     = Nucl.00h-Cyto.00h,
            NuclearVsCytoplasmic_02h     = Nucl.02h-Cyto.02h,
            NuclearVsCytoplasmic_04h     = Nucl.04h-Cyto.04h,

            # Then Nuclear significant from 02h will be compared to Nuclear significant of 00h
            Nuclear_00vs02    = (Nucl.02h-Cyto.02h)-(Nucl.00h-Cyto.00h),
            # Then Nuclear significant from 04h will be compared to Nuclear significant of 00h
            Nuclear_00vs04    = (Nucl.04h-Cyto.04h)-(Nucl.00h-Cyto.00h),

            # Then Cytoplasmic significant from 02h will be compared to Cytoplasmic significant of 00h
            Cytoplasmic_00vs02    = (Cyto.02h-Nucl.02h)-(Cyto.00h-Nucl.00h),
            # Then Cytoplasmic significant from 04h will be compared to Cytoplasmic significant of 00h
            Cytoplasmic_00vs04    = (Cyto.04h-Nucl.04h)-(Cyto.00h-Nucl.00h),

            # Then Nuclear of 00h compared to the mix of nuclear specific from all other timepoints
            Nuclear_00vsTotal = (((Nucl.02h-Cyto.02h)+(Nucl.04h-Cyto.04h))/2)-(Nucl.00h-Cyto.00h),

            # Then Cytoplasmic of 00h compared to the mix of nuclear specific from all other timepoints
            Cytoplasmic_00vsTotal = (((Cyto.02h-Nucl.02h)+(Cyto.04h-Nucl.04h))/2)-(Cyto.00h-Nucl.00h),

            levels=edgerGlmDesign);

I am pretty sure that my conception of contrasts is wrong since results from 'Nuclear_00vsTotal' and 'Cytoplasmic_00vsTotal' are symmetric.

I took inspiration from documentation chapter 3.3.1 where a similar setup is used to compared DrugVsPlacebo.1h, but I probably extrapolated too much from it...


Thank you for your help,
Regards.

[dariober]

It seems to me that Nuclear_00vs02 is equivalent to Nuclear_00vs02, you should get the same results just with opposite sign. The same for Nuclear_00vs04 and Cytoplasmic_00vs04 and the same for Nuclear_00vsTotal and Cytoplasmic_00vsTotal.

I would suggest to run the analysis, pick the top hits and see on the normalized data (i.e. cpm or rpkm) if the result is consistent with what you expect the model to do. Also as a test, add to the raw data some dummy genes that for each contrast should come up as significant or non significant (sort of positive and negative controls), just to make sure the contrasts are doing what you expect.