Learning molecular fingerprints of foods to decode dietary intake

Assessing dietary intake from biological samples provides critical objective insights into nutrition and health. We present a reference-based strategy using untargeted metabolomics to estimate relative dietary composition. The approach learns food-specific molecular ion features first - both annotated and unannotated - via supervised classification and discriminant analysis. These features then guide extraction of corresponding MS1 intensities from unknown samples, enabling proportional, ion-resolved dietary readouts. Tracking these signatures across thousands of public datasets revealed feces, urine, and blood/plasma as optimal biospecimens. Validation with NIST omnivore/vegan stool samples, controlled mouse feeding study, food reintroduction trial in Crohn's disease, and a Mediterranean diet intervention trial confirmed that ion-resolved readouts reflect known intake patterns. In rheumatoid arthritis data, dietary scores obtained from MS/MS signatures correlated with clinical outcomes. To facilitate adoption, we developed an easy-to-use web-based "food readout" app. This method complements traditional diet assessments and advances personalized nutrition and nutritional epidemiology.