
Researchers from P&G's Singapore Innovation Center and Singapore-based Institute of Advanced Intelligence and Computing (IAIC), Agency for Science, Technology and Research (A*STAR), have developed a machine-learning approach that could improve the prediction of odor characteristics from chemical structures, offering a potential new tool for fragrance and flavor development.
The study, published in Flavour and Fragrance Journal, introduces a topology-aware prediction framework that combines conventional molecular descriptors with a mathematical approach called persistent homology. The aim is to capture not only a molecule's chemical composition and structural properties, but also its broader three-dimensional organization, including molecular shape and ring arrangements.
- The research partners developed a machine-learning approach to predict a molecule’s odor characteristics from its chemical structure, potentially supporting fragrance and flavor development.
- The method combines conventional chemical descriptors with a mathematical technique called persistent homology, which captures additional information about molecular shape, three-dimensional organization and ring arrangements.
- Researchers trained and tested the models using 5,213 odorant molecules and 151 expert-labelled odor characteristics, drawing on three curated datasets.
- Among the classification methods tested, random forest performed best. Adding the new structural descriptors improved its micro-F1 score from 0.293 to 0.305, indicating better overall prediction performance.
- The findings suggest that accounting for molecular topology can help machine-learning systems distinguish odor characteristics more accurately, potentially providing fragrance and flavor developers with a stronger tool for exploring and designing scent profiles.
The full details:
The researchers combined molecular descriptors generated by three cheminformatics tools—PaDEL, RDKit and Mordred—with additional descriptors based on persistent entropy, derived from persistent homology. In simple terms, persistent homology is a mathematical method for identifying structural patterns that remain consistent across different scales. Persistent entropy converts this information into fixed-length numerical descriptors that can be used by machine-learning models. These descriptors capture complementary information about molecular topology, including shape and ring organization, that conventional chemical descriptors may not fully represent.
To develop and test the framework, the team assembled a curated dataset from three expert-labelled sources, covering 5,213 odorant molecules and 151 odor characteristics. They evaluated several classification methods, including random forest, decision tree and k-nearest neighbors (KNN). The random forest model delivered the strongest overall performance, achieving a micro-F1 score of 0.305, a macro area under the receiver operating characteristic curve (AUROC) of 0.714 and a micro-AUROC of 0.866.
Importantly, adding the persistent-entropy descriptors improved the random forest model's micro-F1 score from 0.293 to 0.305. Additional feature-permutation analysis confirmed that these descriptors contributed positively to predictions across most of the evaluated odor categories.
The findings suggest that incorporating information about molecular topology can improve how machine-learning systems distinguish between odor characteristics. The researchers say the breadth of odor descriptors covered could support more generalizable prediction and provide a foundation for systematic scent design.









