Entering edit mode
Hi! I am performing differential expression analysis. I used the method apeglm and normal to perform the shrinkage of the log2 Fold Change. I noticed that when I shrinked the LFC with apeglm method all the LFC are almost equal to zero. However, I did not have this pattern with the normal method. I was wondering why apeglm method shrinked the LFC to zero and if this method is appropriate for my data.
Thank you!
Concetta
If I understood well, when I have little differences in my dataset between the two conditions that I am testing, apeglm will overshrink the LFC. So if I have little differences, is it better to use the normal shrinkage method to assess the differences? I attached the MA plot using the MLE method, Normal method and apeglm method. Thank you! Concetta
It's not "overshrinking" really, it's giving output that is justified by the noise level in the data.
I think the
type="normal"is not better than apeglm, as we shown in the apeglm paper. Perhaps you would just want to use ~~p-values~~ adjusted p-values and the MLE (un-shrunken) LFC then.Ok Thank you! So considering that I did not have differential expressed genes with padj < 0.05, I will use the p-value and the MLE (un-shrunken) LFC as you suggested. If I use the un-shrunken LFC can I still use this LFC for visualization and ranking?
Sorry, I meant adjusted p-value above.
I would not recommend using the un-shrunken LFC for ranking. I suppose if the recommended shrinkage methods (apeglm and ashr) are shrinking the LFC to zero, then DESeq2 doesn't have any information for you on optimal ranking of genes for this dataset.