How to interpret result from glmQLFTest in edgeR ?

Question

[Björn]

I am sharing a screenshot after running a command in edgeR

qlf<-glmQLFTest(fit, coef = 2:3)
topTags(qlf, n=5)

![anova like test][1][1]: https://ibb.co/jsDff7 The comparision is between coef 2 and 3 while there is intercept (negative control) as well. How to interpret the pvalue and FDR ? Are these DE genes ? NOTE: This analysis is from edgeRUsersGuide.pdf, section 3.2.6

Answer 1

[russhh]

Are you sure that's exactly as in the edgeR user's guide. I think you've used design <- model.matrix(~ -1 + group) instead of design <- model.matrix(~ group) [strictly, the comparison you've described isn't testing _between_ coefficient 2 and coefficient 3, it's a test to see if either coefficient2 or coefficient3, or some combination thereof accounts for an appreciable amount of the variation for a given probe]

[Björn]

ya, I am sure. edgeRUsersGuide.pdf, section 3.2.6 page 30-31 The design matrix I used is

design<-model.matrix(~0+diagroup)
colnames(design)<-levels(diagroup)
design

The "diagroup" has 5 groups, the first one is negative control (Intercept)

[russhh]

can you have a look at the contents of all of the following design matrices model.matrix(~ -1 + diagroup), model.matrix(~ 0 + diagroup), model.matrix(~ diagroup).

Within section 3.2.6, design is model.matrix(~ group), as defined on p30. There really is a big difference between the design you've just given and the design used in section 3.2.6. Your design fits a separate coefficient for each level of the group; the design model.matrix(~ group) fits a coefficient for the first level and two other coefficients - one for the difference between levels 2&1 and one for levels 3&1

[Björn]

I am not aware of first one you mentioned, however, between

model.matrix(~ 0 + diagroup)
model.matrix(~ diagroup)

, the first model don't keep first group as intercept unlike the second model matrix

[russhh]

then use the second one.

[Björn]

how you would then compare other groups with negative control (intercept) ?

[russhh]

You find genes that are significant for some-combination-of-coefficients using glmQLFTest using all the non-intercept coefficients (using the same code as in your original post glmQLFTest(fit, coef = 2:5)). Then for those that are significant, you run individual A vs B comparisons using glmQLFTest(fit, coef = 2), glmQLFTest(fit, coef = 3), glmQLFTest(fit, coef = 4), glmQLFTest(fit, coef = 5)