Multi-label odor profiling of Osmanthus Oolong tea using graph neural networks: Integrating public databases and PTR-TOF-MS-based aroma compound analysis

Accurate identification of aroma-active compounds is essential for evaluating food quality, sensory characteristics, and authenticity. However, predicting multi-label odor attributes directly from volatile molecular structures remains challenging due to data sparsity and imbalance. In this study, we develop a graph neural network (GNN)-based ensemble framework to support odor profiling in high-aroma tea products. Using two public odor databases (GoodScents and Leffingwell PMP 2001; 4983 molecules, 138 labels), we systematically train and compare multiple GNN architectures (MPNN, GINE, EdgeGAT, PNA) and ensemble strategies. To assess real-world applicability, we analyze four batches of Osmanthus Oolong tea (single and double scented, 2021-2022), identifying 48 volatiles, 16 of which contain odor annotations measured using proton transfer reaction time of flight mass spectrometry (PTR-TOF-MS). Feature-level fusion ensembles achieve the best overall performance, reaching an AUC-ROC of 89.1, precision 66.7, recall 30.2, and an F1-score of 41.6, outperforming individual GNNs (e.g., MPNN 88.7) and traditional machine-learning models such as XGBoost (85.2) and Random Forest (83.3). Incorporating molecular similarity further improves predictions for new compounds. This study demonstrates the potential of integrating odor databases with deep learning to enable data-driven sensory analysis and quality monitoring in flavored foods, offering a scalable solution for aroma evaluation. • Integrated public odor databases with PTR-TOF-MS tea data for validation. • Developed a GNN ensemble for multi-label odor prediction from molecules. • Molecular feature fusion improved accuracy of food aroma prediction. • ML-based odor profiling supports food quality and sensory analysis.