Decoding Smell Through Molecular Machine Learning

Smell links molecular structure to perception through a process that cannot be described by simple chemical rules. Closely related molecules can produce different odor impressions, while structurally distinct compounds may share similar odor characteristics. This complexity makes the prediction of odor from molecular structure a challenging computational problem. In this work, a curated multi-label odor database was constructed by combining two public sources and harmonizing their odor descriptors through an ontology-based framework. The resulting odor database was used to train and evaluate graph-based deep learning models for odor prediction. A directed message passing neural network (D-MPNN) and a discrete flow matching (DFM) model were evaluated, both with and without transfer learning from a larger dataset containing COSMO-RS-derived physicochemical properties. Discrete flow matching achieved the best overall predictive performance reaching a scaffold-split F1_micro of 0.44, although absolute prediction accuracy remained limited by the noise and subjectivity of existing odor labels. Analysis of the learned molecular representations further showed that related odor qualities occupy partially overlapping regions in latent space. Overall, this work demonstrates the potential of graph-based machine learning for molecular odor prediction while highlighting the importance of curated and harmonized odor data for improving structure–odor modeling.