Hello everyone,

I wondered if anyone could help with following problem: I was asked to do SVA for a protein dataset. The goal of the following analysis is to see any inter-individual variation.

For my SVA I have two models (full and null model), where I can include variables of interest and adjusment variables. I have quite some adjusment variables (as gender, age) but struggeling with a "variable of interest". I don't have a vector, as "cancer status", as often mentioned.

My question is: Can I run SVA in any way without a variable of interest? - I assumed not, but was asked to do this. If not, any ideas what could be a "variable of interest", if I'm interested in inter-individual variability?

Thanks in advance

You could use SVA, but it's not clear that you should. The goal for SVA is to identify excess variability in your data and provide ways to adjust for that variability. In this context, excess variability is defined as patterns in the residuals from a linear model, which if the model fits well are not expected to exist. In other words, when you fit a linear model, you are attempting to model the observed variation as a function of known phenotypes. If the coefficients of your model accurately describe changes in the protein abundance, then the residuals for the model will be iid normal variates, with little to no pattern (plotting residuals is one of the main ways to determine if your model fits well and/or fulfills the assumptions underlying the model).

In your case, you apparently have only nuisance variables (differences between subjects that may cause differences in protein abundance but are not of particular interest to you). If that's all you have, then the residuals from a model that contains only nuisance variables contains the variability you are interested in! SVA provides a way to remove that excess variability from your model, but why would you want to remove the thing you are interested in?