Hi Michael and Simon,
I wanted to ask if there was a resent update that could be changing my pvalues in a drastic way?
I ran my analysis on the 18 of August and was getting 2885 DEGs (padj <0.01), today the same .R script (using my same input files) I get 1559 DEGs (padj <0.01). The output stat, pvalues and padj are very different. I also checked with other datasets that I have and got the same different results.
ddsLRT <- DESeq (dds, test ="LRT", reduced= ~Subject)
resLRT <- results(ddsLRT)
resOrdered <- resLRT[order(resLRT$log2FoldChange),]
before (August):
baseMean log2FoldChange lfcSE stat pvalue padj
1.2.23884.m1 6.44702248 -21.40753771 243.210067 22.99579811 1.62E-06 2.81E-05
1.2.17710.m1 6.409085561 -21.39654014 243.2647266 20.83683881 5.00E-06 7.52E-05
1.2.15201.m1 5.287495588 -21.13663198 244.5660123 17.45099162 2.95E-05 NA
1.2.16626.m1 3.685224198 -20.64942877 247.0622647 12.50468092 0.000405934 NA
1.2.23059.m1 1.842612099 -19.71611395 252.0642049 6.687733347 0.009707856 NA
1.2.20081.m1 1.760838826 -19.62249465 252.5823911 8.624799606 0.003316169 NA
1.2.17405.m1 1.522157821 -19.45937386 253.4939475 5.638415871 0.017571076 NA
Today:
baseMean log2FoldChange lfcSE stat pvalue padj
1.2.23884.m1 6.44702248 -21.40611559 243.254081 1.541647987 0.214372672 NA
1.2.17710.m1 6.409085561 -21.39506324 243.3105312 1.09081661 0.296290345 NA
1.2.15201.m1 5.287495588 -21.13518038 244.6113169 1.058806029 0.30348766 NA
1.2.16626.m1 3.685224198 -20.64802728 247.1065218 0.99876432 0.317609691 NA
1.2.23059.m1 1.842612099 -19.71599394 252.0680863 3.763912397 0.052369906 NA
1.2.20081.m1 1.760838826 -19.62096702 252.593144 5.97981623 0.01447051 NA
1.2.17405.m1 1.522157821 -19.45922906 253.4986628 3.12559404 0.077071777 NA
My problem is that my old computer broke so I can not check the DESeq2 version that I was using two months ago, but it should have been up to date.
Do you have any idea of what is going on?
Thanks,
Catalina
hi Catalina,
There was not a recent update, i.e. within the current release of DESeq2 v1.4, that would have changed results at all.
I would guess you were using DESeq2 v1.0 or 1.2 previously. It's hard to say what changed without knowing the version which you ran previously. If you installed R between April 2013 and Oct 2013, it would have likely been DESeq2 version 1.0, whereas if you installed R between Oct 2013 and April 2014, it would have likely been version 1.2. As 1.0 was the first release, there were many changes between 1.0 and 1.2. All the changes which affect end users are recorded in the NEWS file, but these might not be very informative for you.
Note that with test="LRT", there is not shrinkage of log fold changes, so sorting by log2 fold change will not necessarily be the genes with the smallest p-values (whereas with shrinkage of LFC these are typically a similar set to the genes with smallest p-value).
Hi Michael,
Thanks for your reply. But I guess my previous results will still be ok, depending on which genes of interest I will lose between the two different analysis?
As for the LRT test, what I understand from the manual, is to use LRT if I need to define a reduced formula? Which I should if I want to remove the subject effect.
It's not only the LRT that can account for the subject effect. You can also account for the subject effect using the standard steps which produce a Wald test. This is described in the section of the vignette: "Multi-factor designs".