Explainable artificial intelligence models for key-metabolites identification in overweight subjects

Body mass index (BMI) is a common measure used to quantify nutritional status such as overweight condition. Therefore, precise and scalable computational techniques that characterize those phenotypes are essential for clinical care and drug discovery. In this work, we present a pipeline suited to nutritional status evaluation and biomarker identification in the contest of overweight status. We use a multilayer perception network trained on five different datasets with BMI annotations to fulfill this demand. By conducting a comprehensive explainability analysis, we extracted a set of metabolites extremely important from every dataset, which improved our comprehension of the metabolic markers linked to overweight status. Moreover, we identified the relationship between metabolite concentrations and nutritional status and performed an enrichment analysis, which gave us important evidence on the metabolic processes that underlie overweight status.