Olfaction Prediction With Topology‐Aware Machine Learning

ABSTRACT Predicting odour character from molecular structure is a longstanding challenge across the fragrance, flavour, and chemical industries. Conventional cheminformatics descriptors capture many physicochemical and structural properties but may not fully represent the global topological organisation of molecules. Here, we develop a multi‐label odour‐prediction framework that combines descriptors generated using PaDEL, RDKit, and Mordred with persistent‐entropy descriptors derived from persistent homology. These topological descriptors provide stable, fixed‐length summaries of molecular topology, including structural features associated with molecular shape and ring organisation. A curated dataset integrating three expert‐labelled sources contained 5213 odorants and 151 odour characters. Among the classifiers evaluated, random forest performed best, achieving a micro‐F1 score of 0.305 and macro‐ and micro‐AUROC values of 0.714 and 0.866, respectively. Ablation analysis shows that adding persistent entropy increases the micro‐F1 score from 0.293 to 0.305. Moreover, permutation of persistent entropy confirms its positive contribution to most evaluated categories. Together, these results demonstrate that persistent entropy captures complementary global topological information, improves odour‐label discrimination, and strengthens multi‐label prediction. The breadth of odour characters covered offers strong generalisability, providing a useful basis for systematic scent design.