Question

DESeq2 lfcShrink warning use ashr method

0

Entering edit mode

MatthewP • 0

@matthewp-21562

Last seen 3.8 years ago

China

Hello, my lfcShrink function gives warning I don't understand, post on biostar link.

> resLFC_sh2 <- lfcShrink(dds, contrast=c("Treatment", "Sh2", "Control"), type="ashr")
using 'ashr' for LFC shrinkage. If used in published research, please cite:
    Stephens, M. (2016) False discovery rates: a new deal. Biostatistics, 18:2.
    https://doi.org/10.1093/biostatistics/kxw041
Warning message:
In estimate_mixprop(data, g, prior, optmethod = optmethod, control = control,  :
  Optimization failed to converge. Results may be unreliable. Try increasing maxiter and rerunning.

MA plot

res <- results(dds, contrast=c("Treatment", "Sh2", "Control"))
plotMA(res)

plotMa
Plot link

deseq2 • 2.6k views

ADD COMMENT • link 5.2 years ago MatthewP • 0

0

Entering edit mode

Mike, we directed the user here, as we were unsure about this one. I personally had not seen that warning message before. https://www.biostars.org/p/397274/

ADD REPLY • link 5.2 years ago Kevin Blighe ★ 4.0k

score 0 · Answer 1 · 2019-09-06

0

Entering edit mode

Michael Love 43k

@mikelove

Last seen 3 days ago

United States

I added the warning to your post...

Can you post an MA plot of the un-shrunk LFC?

res <- results(dds, contrast=c(...))
plotMA(res, ylim=c(-6,6))

ADD COMMENT • link 5.2 years ago Michael Love 43k

0

Entering edit mode

Ok, added to post now.

ADD REPLY • link 5.2 years ago MatthewP • 0

0

Entering edit mode

It looks like it's not shrinking much, so maybe I'd go with a different method.

You can use:

res <- lfcShrink(dds, coef="treatment_Sh2_vs_Control", type="apeglm")

right?

ADD REPLY • link 5.2 years ago Michael Love 43k

0

Entering edit mode

Yes yes apeglm work.

ADD REPLY • link 5.2 years ago MatthewP • 0

0

Entering edit mode

Hello Michael

I am having a similar problem than that of Matthew, but in my case I was trying to compare several samples on each group.

I have 3 replicates per sample and say 8 unrelated samples, coming from different tissues. Now I want to compare 4 of these samples against the other 4 since they can be grouped in these two groups based on a phenotypic character (but other groupings are also plausible). I'm using the approach given in this link (https://support.bioconductor.org/p/101096/) when they remove the intercept from the design, and I'm having the above error "Optimization failed to converge. Results may be unreliable. Try increasing maxiter and rerunning.".

As I removed the intercept, my resultsNames(dds) now gives a list of the sample names so I can compare, but now using apeglm (and therefore the "coef" parameter) is not suitable since I would not specify a specific comparison conditionAvsB but a sample conditionA.

Should I avoid combining samples the way I did? Is there any major issue in my approach? I've read the DESeq2 vignette but it's mostly guided to the most common cases, which is not my case.

ADD REPLY • link 5.0 years ago cborreda • 0

0

Entering edit mode

You can use apeglm with nearly any design, you just have to relevel the factor and run nbinomWaldTest() to create the new coefficients. This is a little workaround which works perfectly fine, although an extra step of work for the user.

It's discussed in this section:

http://bioconductor.org/packages/release/bioc/vignettes/DESeq2/inst/doc/DESeq2.html#extended-section-on-shrinkage-estimators

ADD REPLY • link 5.0 years ago Michael Love 43k

0

Entering edit mode

Since I want to do group-wise comparisons grouping samples in different ways, I will define different grouping variables and, for each comparison, set a new design and rerun nbinomWaldTest. I guess it is correct (software-wise, I'm aware that the correctness from a biological point of view is on me). Thank you for your help!

ADD REPLY • link 5.0 years ago cborreda • 0