DESeq2: Variable Importance
1
0
Entering edit mode
lolol • 0
@a1ce8025
Last seen 28 days ago
United States

Is there a way to obtain the relative importance of a variable in the design formula of the DESeq2 model? For example, if I have a variable for age in my design formula and I want to know the percentage of variance accounted for by age. I know you can run a likelihood ratio test with a reduced model that does not include the factor to see if it is significant or not. But is there something similar to when you have a multiple linear regression model and calculate the proportion of R-squared associated with each variable? Or any metric that allows you to understand how important each variable is within the design formula.

DESeq2 RNASeq • 299 views
0
Entering edit mode
0
Entering edit mode

Originally posted on Biostars: https://www.biostars.org/p/9474612/#9474612

0
Entering edit mode
@mikelove
Last seen 1 minute ago
United States

We don't offer variance explained because it's a GLM.

Note that in a linear model, the order of the variables dictates the variance explained by each term. So even if we did offer this it would depend on the order of variables, unless you do an added-last test where you have all variables in the model, and then remove each one at a time.

You could do this using the covariates and the VST data if you want the ANOVA-like variance explained, or you could do this within DESeq2 and the GLM, removing one term at a time, and examining the deviance:

http://bioconductor.org/packages/devel/bioc/vignettes/DESeq2/inst/doc/DESeq2.html#access-to-all-calculated-values

0
Entering edit mode

I haven't used it but looks good. I'd recommend to supply the VST data to this package.

0
Entering edit mode

When you say observing the deviance, you mean utilizing the likelihood ratio test with the reduced model (exclude the last variable). Am I understanding correctly?

0
Entering edit mode

Yes, deviance is an output in the mcols, so you could rotate through variables, excluding them in the LRT. Or you could use the other package on VST data.

0
Entering edit mode

Ah yes, I see the deviance now in mcols. Would I use this to calculate the pseudo R-squared (McFaddens), comparing each model with a single factor removed to the intercept only model?

0
Entering edit mode

What you do with the deviance is up to you, I was just pointing out that it is an outputted quantity.