Performance of approaches relying on multidimensional intermediary data to decipher causal relationships between the exposome and health: A simulation study under various causal structures. (August 2021)

• Record Type: Journal Article
• Authors: Cadiou, Solène; Basagaña, Xavier; Gonzalez, Juan R.; Lepeule, Johanna; Vrijheid, Martine; Siroux, Valérie; Slama, Rémy
• Is Part Of: Environment international. Volume 153(2021)
• Publication Date: 2021-08
• Subjects: Exposome -- Variable selection -- Multilayer -- Omics -- Specificity -- Sensitivity -- Reverse causality; BMI Body-Mass-Index -- DAG Directed Acyclic Graph -- DSA Deletion-Substitution-Addition algorithm -- ExWAS Exposome-Wide Association Study -- FDP False-Discovery Proportion -- LASSO Least Absolute Shrinkage and Selection Operator -- MITM Meet-in-the-Middle -- oMITM oriented Meet-in-the-Middle; Environmental protection -- Periodicals; Environmental health -- Periodicals; Environmental monitoring -- Periodicals; Environmental Monitoring -- Periodicals; Environnement -- Protection -- Périodiques; Hygiène du milieu -- Périodiques; Environnement -- Surveillance -- Périodiques; Environmental health; Environmental monitoring; Environmental protection; Periodicals; 333.705
• DOI: 10.1016/j.envint.2021.106509
• Languages: English
• ISSNs: 0160-4120

Abstract

Highlights: Exposome studies suffer from low specificity and sensitivity when looking for causal predictors of an outcome. Intermediary layer (eg methylome) information can help reducing exposome dimension. We proposed an "oriented Meet-in-the-Middle" (oMITM) approach. Simulations demonstrated that oMITM showed improved performance compared to current methods used. oMITM is more robust to reverse causality than methods ignoring intermediary data. Abstract: Challenges in the assessment of the health effects of the exposome, defined as encompassing all environmental exposures from the prenatal period onwards, include a possibly high rate of false positive signals. It might be overcome using data dimension reduction techniques. Data from the biological layers lying between the exposome and its possible health consequences, such as the methylome, may help reducing exposome dimension. We aimed to quantify the performances of approaches relying on the incorporation of an intermediary biological layer to relate the exposome and health, and compare them with agnostic approaches ignoring the intermediary layer. We performed a Monte-Carlo simulation, in which we generated realistic exposome and intermediary layer data by sampling with replacement real data from the Helix exposome project. We generated a Gaussian outcome assuming linear relationships between the three data layers, in 2381 scenarios under five different causal structures, including mediation and reverse causality. We tested 3 agnostic methods considering only the exposome and the health outcome: ExWAS (for Exposome-Wide Association study), DSA, LASSO; and 3 methods relying on an intermediary layer: two implementations of our new oriented Meet-in-the-Middle (oMITM) design, using ExWAS and DSA, and a mediation analysis using ExWAS. Methods' performances were assessed through their sensitivity and FDP (False-Discovery Proportion). The oMITM-based methods generally had lower FDP than the other approaches, possibly at a cost in terms of sensitivity; FDP was in particular lower under a structure of reverse causality and in some mediation scenarios. The oMITM–DSA implementation showed better performances than oMITM–ExWAS, especially in terms of FDP. Among the agnostic approaches, DSA showed the highest performance. Integrating information from intermediary biological layers can help lowering FDP in studies of the exposome health effects; in particular, oMITM seems less sensitive to reverse causality than agnostic exposome-health association studies.