Abstract

Understanding the relationship between DNA methylation (DNAm) and blood pressure (BP) is critical when studying the biological mechanisms underlying cardiovascular disease (CVD). However, the directionality of this association remains unclear: it is not yet known whether epigenetic modifications in DNAm contribute to changes in BP, or whether changes in BP induce alterations in DNAm patterns. Learning this directional asymmetry from observational data is a crucial yet challenging causal discovery problem.

We propose an information theoretic coefficient of asymmetry to quantify and perform inference on said asymmetry within a Generative Exposure Mapping (GEM) framework. This approach models the exposure-outcome relationship through a generative function, Y=g(X), and is extended to noise-perturbed GEMS (Y=g(X)+e) to accommodate outcome contamination. We establish large sample guarantees using data-splitting and cross-fitting techniques. Our method utilizes a tuning-free, data-driven fast Fourier transformation (FFT) based density estimation. We also provide practical diagnostic tools to empirically check key model assumptions.

Using data from n=522 children (including 247 boys and 275 girls) aged 10-18 years in the Early Life Exposures in Mexico to Environmental Toxicants (ELEMENT) study, we examine the directional relationship between BP and DNAm for six candidate genes. Our method reveals a consistent directional influence from blood pressure to DNAm for the FGF5 and HSD11B2 genes. Our framework serves as a practical tool for either generating new scientific hypotheses or formally confirming existing ones about the directionality of relationships in complex biological systems.