Based on prior posts that I've seen on biostars (Building a predictive model by a list of genes), it seems that VST or rlog transformations are best pratice for model building. From what I understand about using variance stabilized transofrmations, they are primarily used for visualizations with distance involved such as PCAs or hierarchical clustering. I'm looking to use glm & multinom to evaluate microRNAs and their relationship with clinical parameters -- but I'm just looking to understand why its important to use vst or rlog.

Is it due to the fact that DESeq2 uses a negative binomial generalized linear model? What would the benefit of using vst/rlog be over using DESeq2 for logistic regression?