Development and cross-tissue validation of a machine learning-based methylation profile score for the cortisol response to stress

Hypothalamic-pituitary-adrenal axis (HPA axis) dysregulation is a risk factor for poor mental and physical health. Animal studies indicate that DNA methylation may regulate the HPA axis, although studies in humans have produced mixed results. A data-driven, machine learning approach may provide new insight.

Using a novel machine learning pipeline, we developed a methylation profile score (MPS) to predict the salivary cortisol response (AUCi) to the Trier Social Stress Test (TSST) from whole blood Illumina Infinium HumanMethylation 450 K BeadChip data (N = 84, mean age = 34, 49% female).

The MPS was associated with the cortisol response in an independent, cross-tissue cohort (N = 53, mean age = 20, 51% female), both before (β = 0.33, 95% CI [0.09, 0.54]) and after a social stressor (β = 0.3, 95% CI [0.09, 0.47]). The effect remained statistically significant after adjusting for age, sex, ethnicity, and smoking, but it became non-significant at the pre-stress (but not post-stress) time point after additional adjustment for cell proportion estimates (pre: β = 0.16, 95% CI [-0.1, 0.47]; post: β = 0.26, 95% CI [0.004, 0.49]).

We developed a DNA methylation-based score that was associated with the cortisol response to stress in a separate, cross-tissue cohort. The predictive signal may depend, at least in part, on differences in cell composition associated with the phenotype. The methods used in this paper can be adapted and expanded upon in future research developing and testing novel epigenetic scores for complex, stress-related phenotypes.