Search
Question: Time course DE analysis in DESeq2 with Wald test
0
12 months ago by
gtechbio0
gtechbio0 wrote:

Hi All,

This is to make sure I do a correct thing (which might be actually wrong).
I have a time series experiment: 0 time point - yeast is growing alone (control), 3,12,24h time point - yeast is growing on human cells, and 24h (control) - yeast is growing alone.

I need to do a time-series DE analysis with DESeq2.

What I did is

colData_yeast<-data.frame(time=factor(x=c(rep("0",2), rep("3",3), rep("12", 3), rep("24",3), rep("24c",2)),levels=c("0","3","12","24", "24c")))
dds_yeast <- DESeqDataSetFromMatrix(yeast, colData=colData_yeast, ~time)
dds_yeast <- DESeq(dds_yeast)

Then I use contrast to compare all time point against 0 timepoint, e.g.

res_yeast_0_3 <- results(dds_yeast, cooksCutoff=FALSE, contrast = c("time","3","0"))
res_yeast_0_12 <- results(dds_yeast, cooksCutoff=FALSE, contrast = c("time","12","0"))

I'd appreciate very much if someone can explain if something is wrong with this approach.

Thank you

EDIT:

One additional question, in case of my data, would it make sense to switch off the LFC shrinkage?

modified 12 months ago by Michael Love19k • written 12 months ago by gtechbio0
1
12 months ago by
Michael Love19k
United States
Michael Love19k wrote:

This makes sense to compare the different time points to time=0.

If you use the latest version of DESeq2, the above code doesn't include LFC shrinkage. You need to do:

res <- lfcShrink(dds, coef=..., res=res)

In order to obtain shrinkage on a coefficient. The shrinkage is useful for visualization and ranking of genes.

If you are using DESeq2 version <= 1.14 however, then the above code does produce LFC shrinkage.

Hi Michael, thank you for your reply.

Could you also please point out what is the conceptual difference between this approach and LRT-based time-course analysis, and whether any of those has advantages or disadvantages over another?

And one more question regarding shrinkage (sorry if a silly one): if you turn off the shrinkage, do you get different results (in terms of p values) compared shrunk LF changes? According to DESeq paper, I'd expect differences between lowly expressed genes. I am asking this because afterwards I will select genes based on p-values, and I am wondering if those two alternative methods in theory give the same results in terms of p-values.

Thank you

Please see the vignette section on LRT for the conceptual difference between Wald testing and LRT. My recommendation is to use the LRT if you want to test multiple coefficients (e.g. differences at every time point after t=0), and Wald for individual coefficients.

With the current version of DESeq2 (v1.16) and going forward, the shrinkage has no effect on the p-values or adjusted p-values. With previous versions it could have small effects, usually not changing the reported DE status of genes, unless a non-zero lfcThreshold was chosen, in which case there could be more significant differences.

Thanks for your time Michael and sorry for coming back to you late.

As far as I understand, with LRT the software will find a gene which became differentially expressed at any time point. In this case we assume, that if gene is changing its expression, this change is always towards one direction - up-regulation for example. But what if lets say at 3h the gene is up-regulated (compared to 0), but in 6h time point it is down-regulated? How the LRT will treat this situation?

Thank you very much!