Hello MAST team,

I would like to use MAST for testing differential expression between cells from two treatment groups. The data is from a nested factorial design, and I would like to know how to best take the structure of the data into account.

Each of the two treatment groups have cells from three different plates, and these plates are unique. Should I fit the data with a mixed model approach with the treatment as a fixed effect and the plate as a nested random effect? Something like this:

lmer.output <- zlm(~ Treatment +(1|Treatment:Plate), data,method='glmer', ebayes=FALSE)

followed by a likelihood ratio test for Treatment?

Or should I go for a fixed effects model with treatment-plate interactions, test the interactions, and if they are not significant refit the model without the plates?

Thank you for your time,
Darius

The random effects model

lmer.output <- zlm(~ Treatment +(1|Treatment:Plate), data,method='glmer', ebayes=FALSE)

seems like the safer bet. In the fixed effects model, you would probably end up with an awkward situation where the interactions are significant in some genes and not others, and I don't have good advice about how to handle that.

There are some other, not terribly well-documented features for the random effects models.
For instance, by default if there is any hint of a convergence issue from lme4, the coefficients and p-values will be set to missing.
You can relax that by setting zlm(..., method = 'glmer', strictConvergence = FALSE).

Let me know how it goes -- in some cases the random effects models seem to converge OK while in other cases, they are problematic.