Abstract
The identification of phenotype-driven network modules in complex, multifluid metabolomics data poses a considerable challenge for statistical analysis and result interpretation. This is the case for phenotypes with only few associations ('sparse' effects), but, in particular, for phenotypes with a large number of metabolite associations ('dense' effects). Herein, we postulate that examining the data at different layers of resolution, from metabolites to pathways, will facilitate the interpretation of modules for both the sparse and the dense cases. We propose an approach for the phenotype-driven identification of modules on multifluid networks based on untargeted metabolomics data of plasma, urine, and saliva samples from the German Study of Health in Pomerania (SHIP-TREND) study. We generated a hierarchical, multifluid map of metabolism covering both metabolite and pathway associations using Gaussian graphical models. First, this map facilitates a fundamental understanding of metabolism within and across fluids for our study, and can serve as a valuable and downloadable resource. Second, based on this map, we then present an algorithm to identify regulated modules that associate with factors such as gender and insulin-like growth factor I (IGF-I) as examples of traits with dense and sparse associations, respectively. We found IGF-I to associate at the rather fine-grained metabolite level, while gender shows well-interpretable associations at pathway level. Our results confirm that a holistic and interpretable view of metabolic changes associated with a phenotype can only be obtained if different layers of metabolic resolution from multiple body fluids are considered.