the dmpFinder function in minfi is great. But the next question is how do I control for co-variates that clearly effect DNA methylation. For instance, it's well known that smoking effects DNA methylation.

Say I'm determining differences in CpG methylation between schizophrenia patients (who are known to be heavy smokers) and healthy controls. If I collect some arbitrary smoking score from all subjects, how can I incorporate and control for smoking score when I calculate differences in CpG methylation between the two groups?

Attempting to answer my own question, it looks like limma can help me. But I don't see any clear workflows/pipelines online that explain how to incorporate and control for co-variates in design matricies created for limma.

Can someone point me to a pipeline that can offer me guidance for what I want to do? if not, any general suggestions or tips would be MUCH appreciated.

NOTE: A cross-package Bioconductor workflow for analysing methylation array data by Maksimovic did not really help as she explained paired analysis with methylation data in limma which is not what I want.

**Im a novice to bioconductor btw. I'd say I'm well versed with minfi, but now that I'm trying to figure out how to build models to answer more complex questions I've hit the wall (despite all the time I've spent reading limma documentations and user guides)

limma can't really help you (although that's what dmpFinder uses 'under the hood'). You can use limma to analyze each CpG, but then you need to do something to aggregate the signal across genomic regions. Which is what bumphunter does. Both bumphunter and lmFit from limma use a design matrix that you specify with model.matrix.

If you want bunches of examples of how to specify a design matrix and how to interpret the coefficients, I would look at the limma User's Guide. Or you could look at anything that is intended to show how to use R to do linear modeling. Or just ?model.matrix or ?formula, the latter of which is pretty good.

If you want more in-depth information, Julian Faraway's book is good as well.

To analyse covariates in limma, just the the covariate in the design matrices. It works just the same as any multiple regression model. It isn't specifically documented because there's hardly anything to say. It just works as one would expect.

As James says, limma can't give a complete methylation analysis by itself because it doesn't aggregate results for CpGs into larger regions. However, if you have a methylation workflow for paired data, adapting it to other linear models with covariates should be reasonably straightward.