Methods/Statistics
Evaluating methods for high-dimensional mediation in metabolomics data Susan Hoffman* Susan Hoffman Donghai Liang Anne dunlop Todd Everson Audrey Gaskins Michele Marcus Ashley Naimi
Background: The metabolome represents a biologically functional measurement of gene-environment interactions and gene transcription. Metabolomics has emerged as a sensitive analytical platform with the potential to offer novel insights into the biological underpinnings mediating exposure-outcome relationships with the appropriate application of statistical methods. This study aimed to test several high-dimensional mediation analysis techniques in simulated metabolomics data to understand the functionality and performance of different methods.
Methods: We simulated data based on the Atlanta African American Maternal Child Cohort with a specific focus on the metabolomic features to understand how changes in the high-dimensional mediator set impact estimator bias, mean-squared error (MSE), and confidence interval coverage. We evaluated high-dimensional mediation analysis methods developed by Zhang, et al. (HIMA) by Gao, et al. (HDMA), as well as a “Meet-in-the-Middle” (MITM) approach proposed by Chadeau-Hyam, et al.
Results: In our preliminary analyses, all evaluated methods demonstrated the ability to detect a mediating effect within a system devoid of correlations between mediators. However, the evaluated methods performed differently with correlations between mediators resulting from the exposure affecting mediator-outcome confounding. Under these conditions involving correlated mediators, MITM appeared to be comparatively less effective than HDMA and HIMA in identifying the correct mediating effect and would often select metabolites that were not among the true mediator set.
Conclusions: This study represents the first comprehensive evaluation of different high-dimensional mediation techniques for metabolomics applications utilizing simulated data. By comparing and evaluating existing mediation methods commonly used in fields like epigenetics, this study bridges the gap between existing methodologies and their potential applicability to metabolomics research.