Question

Using alternative LFC shrinkage methods for computing p-values in DESeq2

0

Entering edit mode

zefrieira • 0

@zefrieira-20933

Last seen 3.9 years ago

In this post, Mike Love explained the reasoning for moving LFC shrinkage into its own step using the lfcShrink function, decoupling testing from shrinkage.

Is there any way to use apeglm-shrinked LFC to compute p-values (similar to using betaPrior=TRUE, but with the new estimator)?

I know that you can get significance levels when you use lfcThreshold with apeglm. Is there a reason that this is only possible when you incorporate lfcThreshold into the null hypothesis?

Thanks!

deseq2 apeglm statistics • 1.7k views

ADD COMMENT • link updated 5.6 years ago by Michael Love 43k • written 5.6 years ago by zefrieira • 0

score 2 · Accepted Answer · 2019-06-02

2

Entering edit mode

Michael Love 43k

@mikelove

Last seen 1 day ago

United States

For apeglm and ashr we compute s-values if you set svalue = TRUE, or for apeglm if you set a threshold. See the apeglm publication for details on the s-value. We found this to make more sense for summarizing over a set of posterior distributions than p-values.

ADD COMMENT • link 5.6 years ago Michael Love 43k

0

Entering edit mode

Thank you very much!

Is there a reason you recommend using a stricter threshold for s-values? Is this something you observed empirically?

Also, are there any plans to unify the interface for getting significance levels based on shrunken LFC? Right now, there are two different interfaces:

If the user wants to get padj values based on the classical shrinkage, they method must use betaPrior=TRUE in the DESeq function.
If the user wants s-values based on ashr or apeglm-shrunken LFCs, they must use betaPrior=FALSE in the DESeq function and then set svalue=TRUE in lfcShrink.

Nevertheless, thanks for your help. From what I've seen so for, it seems like apeglm is giving me more biologically relevant results.

ADD REPLY • link 5.6 years ago zefrieira • 0

1

Entering edit mode

Yes, for example, see here:

https://bioconductor.org/packages/release/bioc/vignettes/apeglm/inst/doc/apeglm.html#specific_coefficients

A very large p-value is 1. A very large s-value is 0.5 (probability of wrong sign is equal to random guessing). In general, they represent fairly distinct conceptual amounts though, so I'd just think about what bound you want on wrong sign LFCs.

Also, along the same lines, no we want be harmonizing these interfaces. betaPrior=TRUE is for all intents deprecated. I didn't remove it altogether for backward compatibility. But it is not recommended (e.g. see the performance of the Normal prior vs apeglm and ashr in the apeglm paper).

Also, the s-value are the more meaningful statistic than p-values for sets of posterior distributions, of which the shrunken LFC represent the mode.

ADD REPLY • link 5.6 years ago Michael Love 43k