Machine learning reveals structural determinants of odor detection thresholds and identifies high-potency food odorants

Odor detection threshold (ODT) quantifies the perceptual potency of volatile compounds and is a key parameter in flavor chemistry. It is widely used to identify key food odorants, calculate odor activity values, and evaluate the contribution of volatiles to overall aroma perception. However, reliable ODT data remain sparse and inconsistent, and are unavailable for many volatiles detected in metabolomics analyses, complicating the identification of key food odorants and the quantitative interpretation of aroma contributions. Here we developed a machine learning framework to predict aqueous ODTs using a curated dataset of 1003 compounds, achieving a test-set R ² of 0.83. External validation with 177 independently measured compounds supported the model’s predictive utility, with 87.0% of predictions falling within a 10-fold deviation and 97.7% within a 100-fold deviation. Structural analysis suggested that aliphatic acids and nitrogen-containing compounds exhibit relatively high ODTs, whereas acyclic sulfur compounds, methoxypyrazines, haloanisoles, and exocyclic esters display low ODTs. Model-guided screening prioritized candidate high-potency odorants, and human sensory evaluation of 12 representative compounds provided preliminary prospective support for the prioritization workflow. These results provide a computational framework for prioritizing potent food odorants and offer data-driven insights into molecular features associated with odor potency in food aroma chemistry.