Question: DESeq2 a lot of genes showing up as differentially expressed that only have 1 sample with any expression
1
gravatar for hsbio
10 weeks ago by
hsbio10
hsbio10 wrote:

Hello,

I've done some differential expression analysis with 3 main groups (3 or 4 biological replicates for each group, 11 total samples). When comparing any two of the groups, there are a number of genes that are statistically significant (in some cases, quite a lot) that have very high log2FoldChanges (~15-35) with high standard errors (~4). Upon further inspection, it seems what is happening with at least some of these genes is that only 1 or 2 replicates (out of 3 or 4) are expressing a gene with the other samples showing 0 expression/counts. Now these samples are from primary tissue, so I'm not surprised that there is quite a bit of variability. However, I'm not sure how to best deal with these genes, I understand that they might still be statistically significant but I'm not sure how to best deal with it.

MA-plot example here: https://imgur.com/a/bNOvVWY

So, I was wondering what the best way the handle this would be. The ways I have thought to do it are:

  1. Just leave it as is.
  2. Pre-filter the data to only include genes that have non-zero counts in at least 2 of the 11 total samples (or perhaps in at least 1 sample of each group)? - This should get rid of some of these cases?
  3. Set some sort of standard error filter (lfcSE), but I have no idea what exact value would be reasonable here or if this is reasonable at all.
  4. Should I use lfcShrink using apeglm? Currently these results are just using results(dds) with a lfcthreshold set. If I use lfcShrink, can I then use the shrunken log2 fold changes while still using the adjusted p values from results(dds)?

Any thoughts on these or any other suggestions would be greatly appreciated.

Thanks.

EDIT: It seems DESeq2 might not be removing any outliers based on the cookscutoff due to how my design and samples are set up (see comment below), however I am still not sure how to best deal with this. Apologies for not mentioning this.

Updated Sample Info:

Samples are from 3 different tissues (3 or 4 replicates per tissue) which are sometimes from the same mice (my design is ~mouse + tissue), however mostly there is only 1 or 2 samples from a mice.

deseq2 • 103 views
ADD COMMENTlink modified 10 weeks ago • written 10 weeks ago by hsbio10
Answer: DESeq2 a lot of genes showing up as differentially expressed that only have 1 sa
1
gravatar for Dario Strbenac
10 weeks ago by
Dario Strbenac1.5k
Australia
Dario Strbenac1.5k wrote:

It seems that the genes which you mention are not being detected as outliers, otherwise they would have p-values of NA. You can set the value of cooksCutoff to a number smaller than 0.99, which will cause more genes to be eliminated from statistical testing. edgeR has a similar approach which uses observation weights to reduce the influence of outliers by giving them less weight, rather than eliminating the gene from the analysis.

ADD COMMENTlink written 10 weeks ago by Dario Strbenac1.5k

Thanks for your response

Looking into this more, I noticed that actually no genes are being filtered as outliers (i.e. based on cooksCutoff) and then I realised that this is due to my design (or at least I think it is). So essentially, my samples are from 3 different tissues which are sometimes from the same mice (my design is ~mouse + tissue), however mostly there is only 1 or 2 samples from a mice. And looking at the cookCutoff, it only works on samples that have at least 3 replicates and because of the mouse sample group, this isn't true for most of them. It seems like this might be the main issue as to why these genes are still being included. Sorry for not mentioning this.

However, I'm not sure what to do here. I could just remove mouse from the design (I can't really see any clear influence of it on a PCA plot) but I'm not sure this is ideal?

Thanks

ADD REPLYlink modified 10 weeks ago • written 10 weeks ago by hsbio10
Answer: DESeq2 a lot of genes showing up as differentially expressed that only have 1 sa
1
gravatar for Michael Love
10 weeks ago by
Michael Love26k
United States
Michael Love26k wrote:

You can remove these genes at the start with

keep <- rowSums(counts(dds) >= 10) >= 3
dds <- dds[keep,]
ADD COMMENTlink written 10 weeks ago by Michael Love26k

Thanks.

This seems to get rid of the majority of those cases.

Just to confirm, if I then still apply lfcShrink with apeglm, using those shrunken log fold changes with the adjusted p values (from results(dds), rather than the calculated s values) is still reasonable for things like visualization/ranking?

ADD REPLYlink modified 10 weeks ago • written 10 weeks ago by hsbio10
1

Yes, this is fine to use shrunken LFC for visualization and ranking, and adj p for FDR sets.

ADD REPLYlink written 10 weeks ago by Michael Love26k
Please log in to add an answer.

Help
Access

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
Powered by Biostar version 16.09
Traffic: 158 users visited in the last hour