Development of a metabolomic risk score for exposure to traffic-related air pollution: A multi-cohort study

To synthesize vast amounts of high-throughput biological information, omics-fields like epigenetics have applied risk scores to develop biomarkers for environmental exposures. Extending the risk score analytic tool to the metabolomic data would be highly beneficial. This research aimed to develop and evaluate metabolomic risk score (metRS) approaches reflecting the biological response to traffic-related air pollution (TRAP) exposure (fine particulate matter, black carbon, and nitrogen dioxide). A simulation study compared three metRS methodologies: elastic net regression, which uses penalized regression to select metabolites, and two variations of thresholding, where a p-value cutoff is used to select metabolites. The methods performance was compared to assess 1) ability to correctly select metabolites associated with daily TRAP and 2) ability of the risk score to predict daily TRAP exposure. Power calculations and false discovery rates (FDR) were calculated for each approach. This metRS was applied to two real cohorts, the Center for Health Discovery and Wellbeing (CHDWB, n = 180) and Environment and Reproductive Health (EARTH, n = 200). In simulations, elastic net regression consistently presented inflated FDR for both high and low effect sizes and across all three sample sizes (n = 200; 500; 1000). Power to detect correct metabolites exceeded 0.8 for all three sample sizes in all three methods. In the real data application assessing associations of metabolomics risk scores and TRAP, associations were largely null. While we did not identify strong associations between the risk scores and TRAP in the real data application, metabolites selected by the risk score approaches were enriched in pathways that are well-known for their association with TRAP. These results demonstrate that certain methodologies to construct metabolomics risk scores are statistically robust and valid; however, standardized metabolic profiling and large sample sizes are required.