Abstract
The umami taste, one of the five basic taste modalities, is traditionally linked to the protein content in food. Developing fast, cost-effective tools to predict a molecule's umami taste is crucial for understanding the molecular basis of flavor and rationalizing food production. However, existing umami predictors are largely limited to amino acid sequences of peptides, restricting their general applicability. In this study, we present VirtuousUmami, a novel machine-learning algorithm capable of predicting umami taste from a query compound's SMILES representation. By utilizing a multi-objective approach, we identified a minimal set of 12 interpretable features that achieve high predictive accuracy. Tested across five food and natural compound databases, this tool paves the way for the rational design of peptide-inspired compounds with specific taste properties.
Introduction
Identifying new umami compounds—the savory "fifth taste"—is a primary goal for the food industry, particularly for creating flavor-enhanced products without excessive sodium or additives. Traditional identification methods, such as sensory panels or high-performance liquid chromatography (HPLC), are resource-intensive and slow.
While computational (in-silico) predictors offer a faster alternative, current models suffer from two major drawbacks: they are typically restricted to peptides and often function as "black boxes," providing little insight into the chemical drivers of taste. VirtuousUmami addresses these gaps by providing a general, interpretable predictor applicable to any chemical structure represented by a SMILES string.
Methodology
The development of VirtuousUmami followed a rigorous multi-objective machine learning pipeline designed for both performance and simplicity:
- Feature Engineering: The team initially extracted 1,613 2D molecular descriptors using the Mordred library. 3D descriptors were excluded to ensure computational efficiency and independence from molecular orientation.
- Dimensionality Reduction: To prevent overfitting, the researchers applied statistical filtering (limma eBayes), narrowing the pool to 324 significant features.
- Multi-Objective Optimization: Using a Pareto-based evolutionary algorithm, the team sought a balance between maximizing accuracy and minimizing the number of features. This resulted in an ensemble of two Support Vector Machines (SVMs) utilizing only 12 key molecular features.
- Interpretability: SHAP (SHapley Additive exPlanations) values were employed to rank feature importance, ensuring the model's decisions are chemically grounded.
Visualizing the Umami Taste Space
A critical finding of the study is that a massive reduction in features does not necessarily lead to a loss in discriminative power. By using t-SNE (t-distributed Stochastic Neighbor Embedding) to project high-dimensional data into two dimensions, the researchers demonstrated that the 12 selected features actually provide a clearer separation between umami and non-umami compounds than the original set of over 1,600 descriptors.
These 12 descriptors include internal autocorrelation (e.g., how mass or volume is distributed) and topological charges, suggesting that the spatial arrangement of atomic properties is a primary driver of umami perception.
Performance and Validation
VirtuousUmami was trained and tested on the UMP442 dataset (140 umami and 302 non-umami peptides). To address class imbalance, random oversampling was applied to the minority umami class. The model demonstrated robust predictive power:
- Accuracy: 95.86% in 10-fold cross-validation and 87.64% on an independent test set.
- ROC AUC: The Area Under the Curve reached 0.91 for the test set, indicating a high degree of sensitivity and specificity.
The most influential feature identified was ATSC1m (centered Moreau-Broto autocorrelation of lag 1 weighted by mass), which highlights the importance of mass distribution across the molecular topology in defining the umami profile.
Discussion and Limitations
While VirtuousUmami represents a significant step toward general taste prediction, the authors acknowledge several limitations. The current training data is heavily biased toward peptides, which narrows the model's Applicability Domain (AD). When screening external databases like FooDB or NPAtlas, only a small fraction of molecules fell within the model's reliable prediction range.
Future work will focus on incorporating more diverse experimental data from non-peptide umami compounds to broaden the model's horizons. To support the broader research community, the authors have made VirtuousUmami available via a user-friendly web interface, enabling the rapid screening of novel flavor candidates.
Summary updated