Hello,

I have some gene expression data for different tissue types with multiple replicates. My species has undergone a historical duplication event and and I am looking to compare the genome copies to identify which one is more highly expressed. Due to the nature of the count data, most parametric models don't seem to be a good fit for my problem. I was looking into the voom paper and am wondering if it is statistically sound to apply linear mixed models to the voom transformed data using external R packages such as lme4. I want to add the tissue types and gene groups as factors and treat the locus (for gene duplicates) as a random effect. I was unsure if limma could accommodate this type of analysis since most examples construct factors for samples, and not groups of genes.

Any feedback is much appreciated. Thanks!

limma-voom computes logcpm values with associated precision weights. You could conduct downstream genewise analyses of the logcpm values using any statistical method that is able to accommodate the precision weights.

Comparing groups of genes is another matter. I do not know of any statistical methods that are able to compare groups of genes because genes are not statistically independent and the nature of the dependence is unknown and almost impossible to estimate. Intergene correlation is extremely common and not just for gene duplicates. Perhaps you could build a linear mixed model for the intergene dependence structure, but that would be very challenging and I haven't seen it done.