A Multimodal Integration Framework for Taste and Odor Prediction and Flavor Molecule Discovery

Abstract

Data-driven molecular property prediction is increasingly used to accelerate molecular discovery and computer-aided design. In flavor science, accurate prediction of molecular taste and odor attributes could facilitate the screening and rational design of candidate flavor molecules. However, this task remains challenging because structure–flavor relationships are complex, single molecular representations may provide incomplete property-relevant information, and the available labels are highly imbalanced. To address these challenges, we implemented and evaluated Uni-Flavor, a multimodal prediction framework embedded in an integrated computational discovery workflow. Uni-Flavor combines three-dimensional molecular representations extracted from a LoRA-fine-tuned Uni-Mol2 model with stability-selected Mordred physicochemical descriptors through a dual-branch fusion architecture incorporating modality-specific projection, multiplicative late fusion, label-aware classification, and imbalance-aware optimization. Independent taste and odor models were trained separately under a common modeling and evaluation protocol, using curated datasets containing 17,202 taste molecules across six categories and 4,983 odor molecules across 138 labels. On held-out test sets, the taste model achieved an accuracy of 95.63% and an AUROC of 97.35%, an AUPRC of 82.05%, and an F1 score of 76.32%, whereas the corresponding values for the odor model were 96.03%, 89.97%, 38.00% and 36.94%. Branch-specific atom-level attribution analysis provided heuristic structural insights into the Uni-Mol2-derived component of the predictions by associating attribution patterns with specific molecular substructures. The trained models were subsequently used to screen approximately 670,000 natural products and coupled with a ScaffoldCAMD strategy for the constrained design of synthetic flavor candidates under chemical-validity, scaffold-assembly, and physicochemical-property constraints, followed by preliminary structural safety filtering. Molecular docking was further used to examine plausible receptor–ligand interactions and provide preliminary receptor-level support for selected candidates. This work presents a common multimodal modeling strategy that integrates three-dimensional molecular representations with physicochemical descriptors and is instantiated independently for multilabel taste and odor prediction. Although developed for flavor-related prediction, evaluation on the SIDER dataset provides preliminary cross-task evidence that the multimodal integration strategy can be applied to another imbalanced multilabel molecular property problem; broader transferability to other property domains remains to be established.