Interpretable Message‐Passing Neural Networks for Molecular Odour Perception Prediction: From Structural Topology to Olfactory Characterisation
- 8 September 2025
- journal article
- Published by Wiley in Flavour and Fragrance Journal
- Vol. 40 (6) , 946-956
Abstract
Molecular odour, as a crucial chemical property, serves as the foundation for flavour modulation in food product development, directly impacting applications ranging from precision design of food flavours and fragrance synthesis to environmental monitoring. However, the structure–activity relationship between molecular characteristics and olfactory perception remains incompletely understood, while traditional experimental approaches face challenges in large‐scale molecular odour characterisation. Recent advances have highlighted the growing importance of artificial intelligence algorithms integrated with chemical data processing techniques for molecular property prediction and rational interpretation of structure‐flavour correlations. This study develops a MPNN‐based fusion model that constructs a multi‐label binary classification framework by integrating molecular topological features with chemical feature maps. Through atomic‐level feature embedding and bond relationship propagation mechanisms, our architecture effectively captures cross‐scale structural patterns from a meticulously curated molecular odour dataset. The model demonstrates robust predictive performance across key food‐related odour dimensions including Odourless, Fruity, Offensive, Oily and Woody (F1‐score: > 0.73, AUC: > 0.82), while maintaining strong generalisation capabilities for low‐abundance labels (e.g., Alcoholic, n = 50: F1‐score = 0.6940, AUC: = 0.8612). The mechanistic analysis revealed that gradient backpropagation can elucidate quantitative associations between specific functional groups and odour perception, providing actionable insights for rational modification of flavour molecules. Our work innovatively integrates multi‐label classification with feature visualisation, overcoming the interpretability limitations inherent in conventional QSAR models for odour prediction. The developed computational framework provides an extensible technical pathway for intelligent fragrance development and odour digitisation engineering, establishing a novel algorithmic paradigm for next‐generation food sensory innovation.Keywords
This publication has 37 references indexed in Scilit:
- Progress in flavor research in food: Flavor chemistry in food quality, safety, and sensory propertiesFood Chemistry: X, 2024
- Simultaneous strengthening and toughening lignin/cellulose nanofibril composite films: Effects from flexible hydrogen bondsChemical Engineering Journal, 2022
- Deep Learning-Based Increment Theory for Formation Enthalpy PredictionsThe Journal of Physical Chemistry A, 2022
- Comprehensive strategies of machine-learning-based quantitative structure-activity relationship modelsiScience, 2021
- Graph-Based Approaches for Predicting Solvation Energy in Multiple Solvents: Open Datasets and Machine Learning ModelsThe Journal of Physical Chemistry A, 2021
- SMILES to Smell: Decoding the Structure–Odor Relationship of Chemical Compounds Using the Deep Neural Network ApproachJournal of Chemical Information and Modeling, 2021
- Applying medicinal chemistry strategies to understand odorant discriminationNature Communications, 2016
- Relationships between volatile compounds and sensory characteristics in virgin olive oil by analytical and chemometric approachesJournal of the Science of Food and Agriculture, 2015
- Olfactory detectability of homologous n-alkylbenzenes as reflected by concentration-detection functions in humansNeuroscience, 2009
- Effects of functional group position on spatial representations of aliphatic odorants in the rat olfactory bulbJournal of Comparative Neurology, 2005