Epigenetic prediction of complex traits and death

Background Genome-wide DNA methylation (DNAm) profiling has allowed for the development of molecular predictors for a multitude of traits and diseases. Such predictors may be more accurate than the self-reported phenotypes, and could have clinical applications. Here, penalised regression models were used to develop DNAm predictors for body mass index (BMI), smoking status, alcohol consumption, and educational attainment in a cohort of 5,100 individuals. Using an independent test cohort comprising 906 individuals, the proportion of phenotypic variance explained in each trait was examined for DNAm-based and genetic predictors. Receiver operator characteristic curves were generated to investigate the predictive performance of DNAm-based predictors, using dichotomised phenotypes. The relationship between DNAm scores and all-cause mortality (n = 214 events) was assessed via Cox proportional-hazards models. Results The DNAm-based predictors explained different proportions of the phenotypic variance for BMI (12%), smoking (60%), alcohol consumption (12%) and education (3%). The combined genetic and DNAm predictors explained 20% of the variance in BMI, 61% in smoking, 13% in alcohol consumption, and 6% in education. DNAm predictors for smoking, alcohol, and education but not BMI predicted mortality in univariate models. The predictors showed moderate discrimination of obesity (AUC=0.67) and alcohol consumption (AUC=0.75), and excellent discrimination of current smoking status (AUC=0.98). There was poorer discrimination of college-educated individuals (AUC=0.59). Conclusions DNAm predictors correlate with lifestyle factors that are associated with health and mortality. They may supplement DNAm-based predictors of age to identify the lifestyle profiles of individuals and predict disease risk. List of abbreviations DNAm DNA methylation BMI Body mass index AUC Area under the curve CpG Cytosine phosphate Guanine dinucleotide EWAS Epigenome-wide association study GS:SFHS Generation Scotland: The Scottish family health study LBC1936 Lothian birth cohort 1936 LASSO Least absolute shrinkage and selector operator HR Hazard ratio CI Confidence interval STRADL Stratifying resilience and depression longitudinally