Identification, analysis, and interpretation of a human serum metabolomics causal network in an observational study

Untargeted metabolomics, measurement of large numbers of metabolites irrespective of their chemical or biologic characteristics, has proven useful for identifying novel biomarkers of health and disease. Of particular importance is the analysis of networks of metabolites, as opposed to the level of an individual metabolite. The aim of this study is to achieve causal inference among serum metabolites in an observational setting. A metabolomics causal network is identified using the genome granularity directed acyclic graph (GDAG) algorithm where information across the genome in a deeper level of granularity is extracted to create strong instrumental variables and identify causal relationships among metabolites in an upper level of granularity. Information from 1,034,945 genetic variants distributed across the genome was used to identify a metabolomics causal network among 122 serum metabolites. We introduce individual properties within the network, such as strength of a metabolite. Based on these properties, hypothesized targets for intervention and prediction are identified. Four nodes corresponding to the metabolites leucine, arichidonoyl-glycerophosphocholine, N-acyelyalanine, and glutarylcarnitine had high impact on the entire network by virtue of having multiple arrows pointing out, which propagated long distances. Five modules, largely corresponding to functional metabolite categories (e.g. amino acids), were identified over the network and module boundaries were determined using directionality and causal effect sizes. Two families, each consists of a triangular motif identified in the network had essential roles in the network by virtue of influencing a large number of other nodes. We discuss causal effect measurement while confounders and mediators are identified graphically.