1
0
Entering edit mode
Yuan Hao ▴ 240
@yuan-hao-3658
Last seen 8.0 years ago
United States

Hi Michael,

I used DESeq2 version 1.4 on some RNSeq data four month ago, and repeat the analysis within version 1.5 last week on the same dataset. I got pretty much the same DE gene lists (>90% overlaps), only the fold changes looked much shrunk from version 1.5. We have some rtPCR data correlated better to the old fold change ranges though. I am aware of you’d changed the beta prior variance estimation. Could you provide more details about it? Thanks a million in advance!

Cheers,

Yuan

deseq • 1.4k views
1
Entering edit mode
@mikelove
Last seen 22 hours ago
United States

hi Yuan,

Information about changes is typically in the NEWS file of a package, which can be queried with the news() function:

news(Version >= "1.4.5", package="DESeq2")

My note in the NEWS file doesn't give too much detail however. I'll make a note to add more detail in the nbinomWaldTest man page.

"Adding an alternate method for beta prior variance calculation
in nbinomWaldTest. This helps to produce more robust prior
variance estimates when many genes have small counts and highly
variable MLE log fold changes."

The change was to add a weighting to the contribution of MLE fold changes to the prior variance. Specifically they are weighted by the expected variance of log fold changes using the fitted dispersion trend. So the highly variable MLE fold changes from low count genes barely contribute. It's tricky to obtain the perfect amount of shrinkage to optimize the bias-variance tradeoff, so it's certainly possible that the older method was better for a certain set of genes for a certain dataset.