RNA-seq DESeq2 : p-values and venn plots in same analysis

Question

[BioHazzard]

I am doing differential expression analysis. I am comparing two different experiments, each experiment consisting of two treatments and their respective controls in duplicate.

I used DESeq2 to generate a distinct results object for each of the 4 control/treatment pairs and am doing downstream analysis on genes with the adjusted p-value below 0.01.

My question regards the difference between considering genes differentially expressed based on the p-value, which is continuous and comparing the result with a heatmap. Again, p-value thresholds are taken from the DESeq2 results object generated for each of the conditions.

I will illustrate this with two images. These images take into consideration only two of the 4 conditions.

The venn diagram looks like this:

So in each condition a certain number of genes were differentially expressed and the overlaps between the two conditions are shown. In this example, in condition A there are 275 genes that are only differentially expressed in that condition.

However, when I create a heatmap of those genes, which should be exclusively differentially expressed in condition A, I observe that there is also an obvious difference in condition B, even if less strong. Note that the columns in the heatmap are ordered:

        CTR  CTR  CTR CTR TREAT TREAT TREAT TREAT
         A    A    B   B    A     A     B     B

The heatmap tells a different story than the venn diagram. While simply using the p threshold I can define genes as being uniquely differentially expressed in one condition only, the heatmap makes conditions A and B look much more similar, as also shown by the clustering.

Any tips or insight would be greatly appreciated.

[caggtaagtat]

Hi, I dont know if its the hospital firewall, but the images are not visible to me.

[BioHazzard]

Can you recommend me an image hosting service that you know you can see?

[caggtaagtat]

No, I think I have to apply for a change in my firewall. Just wanted make sure its because of me

How did you create your DESeqDataSet ?

Answer 1

[devbt15]

As we cannot see the sample names it is hard to comment but I would presume that the order is same as in the text provided above. As such a heat map command would portray the absolute values from DESeq2 and here we can see that it is similar in both A and B in control and treatment respectively (It will not consider the p-value while doing so, which you considered on the other hand while calculating your DEGs in the Venn). This would mean that there is narrow expression difference between A and B samples (treatment vs. control). I would suggest you plot log2Fold change (treatment vs. control) for A and B (so containing 2 columns only), to see a better difference and also scale the data before plotting (so the heatmap scale goes from -1 to +1). Regards.

[BioHazzard]

Thanks for the reply. My question is less about how to graph the data but more about how to interpret the data.

If I am trying to make statements such as "275 genes showed modified expression only in condition A", they the venn diagram could lead me to make such a statement. However, I am reluctant to make such a statement because when I look at the heatmap, it tells me that those genes are clearly also regulated. Evidently, they are regulated, but the magnitude of differential expression is smaller, so that their p-values are above the threshold.

It looks to me as if venn diagrams are not very good for DE analysis.

Is there another way to analyse similarity between the conditions?

[caggtaagtat]

Can you maybe group you data in CTR_A, CTR_B, TREAT_A, TREAT_B and than do DGE between condition CTR_A and CTR_B using one of them as the reference? Or are the experiments to different?