Construction of a machine learning-based predictive model for odor-driven behavioral preference of odorant molecules.

How the chemical structures of odorant molecules determine biologically oriented behaviors remains a central challenge at the intersection of food informatics and biology. To identify structural features of odorant molecules that are associated with mouse first-encounter approach-avoidance responses, this study used a published human flavor semantic descriptor database only as a stratification tool to select a semantically diverse set of odorants; the behavioral labels came from the mice's own first-encounter approach-avoidance responses. By integrating multidimensional molecular representations with machine learning algorithms, we established a predictive method for odor-driven behavioral preference. The results showed that, among the prespecified model-representation pipelines evaluated in this study, the Logistic Regression model based on structural keys and physicochemical descriptors achieved the highest mean F1-score (0.812 ± 0.016). Model interpretability analysis indicated that ester groups, aromatic rings, ethers, and branched carbon motifs were more likely to be associated with approach behavior, whereas thioethers, sulfur-containing heterocycles, disulfides, and certain carbonyl/heteroatom environments were more likely to be associated with avoidance behavior. Behavioral experiments using structural analogs provided supporting evidence for an association between local structural modification and shifts in first-encounter approach-avoidance behavior in mice. This study establishes an interpretable olfactory prediction methodology. This methodology provides an interpretable computational framework to support the screening and prioritization of candidate odorant molecules and may inform subsequent human sensory evaluation and product-level validation.