By what means did you attain your methylome data? I ask because I know of one method for doing such analysis off the top of my head but it is for methylation data attained from Infinium 450K arrays and expression data. The publication describing it can be found here: Article A systems-level integrative framework for genome-wide DNA me...

Yes, my data comes from Infinium 450K. Many thanks Pawel! I've not come across this publication yet.

Although the algorithm doesn't explicitly calculate what I am after, the paper provided me with some valuable clues.

I gotta say I'm extremely skeptical of your proposed approach for the following reasons:

1. The causality between DNA methylation and gene expression is not as black and white as people often take it to be. Does DNA methylation actually cause expression repression, or do repressed genes tend to get methylated? Is causality always in the same direction, or does it depend on the context?

2. Even if we assume DNA methylation causes repression of gene expression, it is extremely inconsistent as a predictor of gene expression. As an example of this, look at Fig. 4 of this paper Article Irizarry RA, Ladd-Acosta C, Wen B, Wu Z, Montano C, Onyango ...

3. Methylation across a gene/region can be highly correlated. If you include correlated methylation signals from a gene as separate terms in the same linear model, you will kill the effect size and significance of each correlated term. Combine this problem with the multiple testing burden of looking at all genes, and I don't think anything could be even close to genome-wide significant. It would probably be better to test one methylation signal per model, but if you do this per signal, the multiple testing correction gets even worse, and if you do this with only one signal per gene, you are cherry picking and hoping the methylation signal you choose is one that's relevant to the expression of its cognate gene.

Thanks Jason, but I don't understand what the take-home message is for me.

The authors of the publication posted by Pawel take the average methylation value near the TSS and demonstrate a considerable amount of correlation. Since methylation is (anti)correlated with expression, then it is a factor that may confound other analyses, right?

Perhaps I'm misunderstanding what you ultimately want to do. I interpreted your post as saying you want to generate linear models for predicting gene expression level from DNA methylation betas. My point was that, while there is a significant anti-correlation between methylation near the TSS and gene expression across all genes, for any particular gene, all bets are off. As you say, methylation is not the only source of transcriptional variation, and given an r2 of less than 0.3 (as reported in the Jiao paper Pawel posted), it on average explains a fairly small proportion of the variance in gene expression. I was trying to emphasize that on its own methylation is a poor predictor of expression levels, and even when there is an association, you should be careful about assuming the direction of causality.

However, if you are not trying to predict expression levels from just methylation betas, but are rather interested in a holistic consideration of the sources of transcriptional variance, then I think including methylation in a model along with other factors (i.e. to account for potential confounding with other factors) seems reasonable. Does that make sense?