DNA methylation (DNAm) in neurological tissue is often difficult or impossible to use for observational studies, as sample collection must usually occur post-mortem. Studies interested in DNAm changes in brain related to disease outcomes or exposures commonly rely on DNAm measured in less invasive surrogate tissue, like blood, which typically have poorly characterised relationships with the target tissue DNAm levels of interest.

Early analysis has shown that some correlation exists between blood DNAm levels and DNAm at four different brain regions,1 but so far such work has been limited to comparisons at individual CpG sites. More recently, analyses developing multi-CpG prediction models have been able to serve as proxies for other types of molecular phenotypes, such as circulating protein levels in a variety of contexts. Thus, robustly trained, multi-CpG site blood DNAm models may be able to improve the variance explained of brain DNAm levels, especially in comparison to single CpGs. However, availability of appropriate paired blood-brain DNAm datasets required for model training has been limited to date.

We propose to develop models of brain DNAm using blood DNAm using an existing non-ALSPAC dataset and then apply the best performing models in ALSPAC DNAm to determine if they are associated with phenotypes and exposures relavant to the brain.