When using a negative binomial generalized linear model to identify differentially expressed genes, how does the model account for biological replicates that show higher variance than the mean?

A negative binomial generalized linear model accounts for overdispersion, which is the phenomenon where biological replicates show higher variance than their mean, by introducing a specific dispersion parameter. In a standard Poisson model, which is often used for count data, the variance is assumed to be exactly equal to the mean. Biological gene expression data, however, typically exhibits variance that increases quadratically with the mean, exceeding Poisson ex....