DE between different groups : post-hoc tests?

[fabien-tell]

Hello,
I'm analyzing single cell RT-PCR experiments. For each cell 57 gene expressions are measured and expressed as "et" as required. I want to know if there is DE between 3 locations (factor= location : ( A, B, C) in the cData file)
In code

sca <- FromMatrix(genes_W_T, cData, fData)

zlm.output <- zlm(~ locations, sca)
show(zlm.output)
zlm.lr <- lrTest(zlm.output, 'location')
pvalue <- ggplot(melt(zlm.lr[,,'Pr(>Chisq)']), aes(x=primerid, y=-log10(value)))+
    geom_bar(stat='identity')+facet_wrap(~test.type) + coord_flip()
print(pvalue)

Actually, I only get three p_values (continuous, discrete, hurdle) by gene. I was expecting to obtain at least two series (A vs B) and (A vs C ) groups.
I may have misunderstood. The output of your zlm model may yields to an overall difference between groups ( ANOVA does ) and then we have to do post-hoc tests when the number of groups is greater than 2? If it is the case how to do it with Mast and are FDR correction available?

Thanks a lot

[amcdavid]

lrTest(zlm.output, 'location') drops the entire location term from the model, so a 2 degree of freedom test. To test an individual coefficient is zero you can do something like lrTest(zlm.output, CoefficientHypothesis('locationA')

[fabien-tell]

Thank you.
I had not understood well.
It means that with different factor levels (A, B, C), the test has to be done several times ? :

lrTest(zlm.output, CoefficientHypothesis('locationA')
lrTest(zlm.output, CoefficientHypothesis('locationB')
lrTest(zlm.output, CoefficientHypothesis('locationC')

In this case, what do the p-values for each gene represent ?
With only two levels (e.g. location = A B) I had understood that they represented the difference between A and B location for each gene but now with (A, B, C) I 'm confused. Each above code lines will give me a p-values object with different metrics (lambda, df, p_values) for continuous, discrete and hurdle models but I do not understand it well which comparison is made.

My goal is to test whether gene expression differs between A and B, A and C and B and C. I could use a non-parametric ANOVA followed by post-hoc but this approach looks better because as you may know with single cell PCR, distribution are characterized by a lot of zeros values.

Thanks for your help and patience

[gfinak]

@fabien-tell okay to close this issue? Do you still have questions?