cannot use expanded matrix in DESeq2
1
0
Entering edit mode
KELVINLEE • 0
@kelvinlee-9111
Last seen 9.0 years ago
Singapore

Hi, 

it seems like with the new version of DESeq2 (Version 1.10.0) is always using the standard model matrix as default rather than the expanded model matrix for design with interaction. I tried to switch it back to using expanded model matrix by doing the following, but got an error. Please help.

my experiment has two factors (Type with 2 levels and Day with 3 levels)

experimental design I use is ~Type + Day + Type:Day

> dds <- DESeq(dds, modelMatrixType = "expanded")
estimating size factors
estimating dispersions
gene-wise dispersion estimates
mean-dispersion relationship
final dispersion estimates
fitting model and testing
Error in nbinomWaldTest(object, betaPrior = betaPrior, quiet = quiet,  : 
  expanded model matrices require a beta prior
>

deseq2 • 1.8k views
ADD COMMENT
0
Entering edit mode
@mikelove
Last seen 11 hours ago
United States

hi Kevin,

See a related thread here:

Contrast All Interaction Effects in DESeq2 in current version (1.10.1)

In version 1.10, we no longer support LFC shrinkage (and so expanded matrices) for designs with interactions, which should greatly simplify the interpretation of results for our users. We had the expanded model matrices for interaction terms for 3 releases (1.4, 1.6, and 1.8) and by the last release, I had already noticed that it was far too confusing for users to put together what the terms meant, and why some terms were undergoing shrinkage and others not. So I'm convinced that this was the right choice.

And not to worry, any comparison that you had before can be made using the standard model matrices.

ADD COMMENT
0
Entering edit mode

Hi Michael,

thanks for your fast response. I have taken a look at the thread and I must also agree that it is a pity to take away the extended model matrix. I find it very useful and convenient too.

I still have one question. With the design of of ~Type + Day + Type:Day and the extended model matrix, I can use contrast=c("Type", "mixed", "single") to call for the overall main effect of Type across all levels of Day. Now, with the standard model, what should I contrast or call in order to have the same comparison? I know that using the same contrast=c("Type", "mixed", "single") will only give me the main effect of Type for the reference level of Day with the standard model matrix. Do note that I hope to keep the experimental design identical if possible. Thank you.

ADD REPLY
0
Entering edit mode

hi,

You'll just have to trust me that the latest version is in my opinion the best implementation. The prior calculation for non-interaction terms was straightforward to implement. But the old version of dealing with interactions and priors made too many arbitrary algorithmic choices, and I'm happy to cut out that code for something simpler. 

The way to test for an overall, or average, effect of type across all levels of day is to use the same design as in the link above. That would be here ~day + day:type. Then you will have a type effect for each day in the resultsNames. You can just average these to obtain the average type effect. This would look like results(dds, contrast=list(c("day1typeT","day2typeT","day3typeT")), listValues=1/3). You can contrast these terms to test if the type effect is different on day 3 vs day 2, etc.

ADD REPLY
0
Entering edit mode

Hi Michael,

just one more question. If I use a different design for each run, will I get different results with the same data set? I ask this question because I would really like to use ~type + day + type:day to get some outputs (basically the vsd transformed data) to run PERMANOVA. Eventually, I will have no significant different for interaction (type:day) and day. I will then just look at type by re-running ~type. Do you think this is appropriate?

ADD REPLY
0
Entering edit mode

Yes, different design will give different coefficients and test statistics, and these can be drastically different. You need to choose a design which makes sense for the experiment and go with that. You should use the same design for visualizing the dataset (VST, and here i would recommend blind=FALSE...see vignette for details) and for significance testing (here we recommend testing with DESeq() but it's ultimately up to you in the end what you do).

ADD REPLY

Login before adding your answer.

Traffic: 760 users visited in the last hour
Help About
FAQ
Access RSS
API
Stats

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.

Powered by the version 2.3.6