the dispersion estimation of edgeR and DESeq2

Question

[tommy]

Hello, I am currently studying the dispersion estimation methods used in edgeR and DESeq2.And I wonder if someone can help me correct my understanding.

1. The edgeR and DESeq2 both use the EB method to borrow information from genes in the dataset.
2. In edgeR, the common dispersion is estimated from all the genes in the dataset and acts as a baseline or reference point at first. Then the gene-specific dispersion estimates are shrunken toward this common dispersion.
3. In DESeq2, the specific gene-wise dispersion is estimated. Subsequently, the prior distribution is fitted representing the overall trend of the dispersion-mean relationship. After that, the EB method is used as a shrinkage method.

Based on that, I wonder in the procedure of dispersion estimation, the main difference between these two methods is the order? Since based on my data, they give quite different results. in edgeR, the raw dispersion is estimated using estimateGLMTagwiseDisp(d2,design.mat, prior.df=0). The red dots represent the final results.

The dispersion estimation varies significantly when expression levels are low.

And I wonder what the line at the left bottom corner of both plots means, with the lowest raw dispersion values?

Thanks for the help.