Predicting multiple taste sensations with a multiobjective machine learning method

The study addresses the challenge of predicting multiple taste sensations of chemical compounds from their molecular structures. While taste and smell guide food selection via five basic tastes (sweet, bitter, sour, salty, umami), existing computational tools predominantly focus on single tastes (especially sweet and bitter) and binary classifiers. Real-world foods often exhibit complex blends of tastes, yet simultaneous multi-taste prediction remains underexplored. The authors aim to develop a multi-class machine learning predictor, VirtuousMultiTaste, that classifies compounds into bitter, sweet, umami, or other, leveraging physicochemical features to improve understanding of taste–structure relationships and to support applications in food design and multi-sensory perception analysis.

Prior work includes numerous ML-based tools for predicting individual tastes: for bitterness (e.g., BitterX, BitterPredict, e-Bitter, iBitter-SCM, BERT4Bitter, iBitter-Fuse, QSTR-based methods), for sweetness (e.g., e-Sweet, Predisweet, BoostSweet, BitterSweetForest, BitterSweet, VirtuousSweetBitter), and several umami predictors (iUmami-SCM, UMPred-FRL, VirtuousUmami, Umami-MRNN, Umami-BERT). Early models used MLR and SVM for binary classification, later outperformed by tree-based methods (RF, AdaBoost) and neural networks, which also support multi-class settings given sufficient data. Multiclass and multi-label methods have seen applications in food/agriculture (e.g., tea sample classification, wine aging discrimination, raw food classification), but comprehensive multi-taste prediction and intensity estimation remain limited. This gap motivates a unified multi-class predictor that can handle multiple taste sensations concurrently.

Data curation: Public datasets with verified tastes were aggregated across nine labels (sweet, bitter, non-sweet, umami, tasteless, sour, salty, multitaste, other) from VirtuousSweetBitter resources, UMP442 (for umami/non-umami), and ChemTastesDB. Due to scarce data for sour (38) and salty (12), these were not modeled separately; ‘multitaste’ was excluded. Classes used were Sweet, Bitter, Umami, and Other (the latter aggregating tasteless and miscellaneous tastes). Initial dataset: 6309 compounds (2741 sweet, 2549 bitter, 238 umami, 781 other) with SMILES. Structure standardization and filtering: ChEMBL Structure Pipeline was applied for checking, standardization, and parent structure generation; incorrect SMILES and duplicates removed, yielding 4717 compounds (1904 sweet, 1937 bitter, 227 umami, 649 other). Train/test split: Training subset included 360 sweet, 360 bitter, 227 umami, 360 other; the umami class was oversampled by 133 using an AdaBoost-based procedure to balance classes (to 360 per class). Test set (left out): 3377 compounds comprising 1544 sweet, 1577 bitter, 289 other (no umami in test). Feature engineering: 1613 2D Mordred descriptors were computed per compound. Features with >30% missing values were removed; remaining missing values were imputed via kNN-impute (k=20). Features were normalized to [0,1]. Dimensionality reduction and statistics: Normality assessed by Shapiro–Wilk; due to non-normality, Kruskal–Wallis tests were used with Benjamini–Hochberg FDR correction (q<0.05), yielding 1306 significant descriptors. PCA on all and on significant features illustrated limited linear separability; pairwise taste-vs-rest comparisons used Mann–Whitney tests. Top differentiating features per taste were visualized (supplementary). Model construction: An ensemble dimensionality reduction and model selection pipeline used a heuristic multi-objective Pareto-based evolutionary optimization to (a) select an optimal feature subset, (b) choose classifier (SVM vs RF), and (c) tune hyperparameters (C, gamma for SVM; number of trees for RF). Objectives and weights: Selected Features Minimization (1), ACC (10), F1 (10), F2 (1), Precision (1), Recall (10), AUC (1), Number of SVs/Trees Minimization (1), Manhattan Distance (1). The evolutionary algorithm used population size 100, up to 200 generations; 10 independent runs; convergence observed ~50 generations. Stratified 10-fold cross-validation on training data guided selection. Class imbalance handling: Adaptive Boosting (AdaBoost) was used as a pre-processing step during CV to increase weights for minority class (umami) and generate synthetic copies to balance classes across folds. Selected model: Random Forest outperformed SVM in multi-objective optimization. A final RF model with 95 trees and 15 selected descriptors (from ATS autocorrelation and related Mordred families) was chosen for best trade-off of performance and simplicity. Explainability: SHAP (TreeExplainer) provided per-class feature attributions. The 15 final features: ATSCOc, ATSCOse, AATS0i, ATSC1p, AATSC2se, AATSC0m, AATSC1Z, AATSC2are, AATSC1pe, SpDiam_A, ATSC1c, ATSC1se, ATSC1Z, ATSC1m, ATSC4s. Correlations among these features were analyzed. External screening: After standardization and descriptor computation, the trained model screened five external databases (FooDB, FlavorDB, PhenolExplorer, Natural Product Atlas, PhytoHub) to estimate taste distributions. A web platform (Ionic front end; Flask back end; REST API) enables predictions from SMILES, FASTA, InChI, SMARTS, or PubChem name; outputs include SMILES, 2D depictions, class predictions, and downloadable descriptor sets. Coffee and chocolate compositions (from FooDB) were also analyzed via the platform.

• The multi-objective optimization selected Random Forest (95 trees) with 15 features as the best-performing and most parsimonious model.
• Cross-validation (10-fold, training set): ACC 76.54% ± 1.0; F1 76.58% ± 1.0; F2 76.61% ± 1.01; Precision 76.92% ± 1.05; Recall 76.64% ± 1.0; AUC 0.92 ± 0.02. ROC AUC per class (CV): Bitter 0.92; Sweet 0.92; Other 0.90; Umami 1.00; micro/macro average 0.94.
• Test set (n=3377; 1577 bitter, 1544 sweet, 289 other): ACC 71.76%; F1 74.32%; F2 73.10%; Precision 78.98%; Recall 71.76%; AUC 0.87. ROC AUC per class (test): Bitter 0.89; Sweet 0.86; Other 0.86; micro/macro average 0.87. Reported per-class recalls (test): Bitter 74.50%; Sweet 69.23%; Other 70.24%.
• Feature importance: 15 top Mordred descriptors are predominantly ATS autocorrelation descriptors weighted by electronegativity, charge, mass, polarizability, ionization potential, atomic number, intrinsic state, plus SpDiam_A, aligning with physical properties relevant to tastant–receptor interactions.
• External database screening predictions: • FooDB (n=69,309): Bitter 14,693; Sweet 5,375; Umami 3,149; Other 46,092. • FlavorDB natural ligands (n=2,599): Bitter 778; Sweet 1,661; Umami 29; Other 131. • PhenolExplorer (n=489): Bitter 365; Sweet 23; Umami 9; Other 92. • Natural Product Atlas (n=32,491): Bitter 26,653; Sweet 2,019; Umami 1,880; Other 1,939. • PhytoHub (n=1,746): Bitter 1,213; Sweet 228; Umami 62; Other 243.
• Food examples (from FooDB compositions via platform): Coffee predicted compounds—Bitter 130; Sweet 44; Umami 4; Other 14. Chocolate—Bitter 96; Sweet 33; Umami 4; Other 13.
• Benchmarking: Against RF, XGBoost, and SVM pipelines with mRMR, VirtuousMultiTaste achieved superior metrics across the board. On an external set of 869 compounds (409 bitter, 460 sweet) excluding training overlaps, VirtuousMultiTaste achieved ~83% across metrics for bitter, outperforming VirtuousSweetBitter (~80%) and BitterSweet (~77%); for sweet, performance was intermediate but satisfactory. For umami, cross-validation comparisons with VirtuousUmami showed similar ACC (~96%) and AUC (>96%). Three recently validated umami peptides (FR-9, FE-5, EK-5) were correctly predicted as umami.

The study fills a notable gap by enabling simultaneous prediction of multiple tastes (bitter, sweet, umami vs other) using a unified, interpretable, and efficient model. By combining statistical filtering, 2D Mordred descriptors, and multi-objective evolutionary optimization, the approach balances performance with simplicity (15 features, 95-tree RF). Robust cross-validation and stable performance across similarity quartiles suggest a broad applicability domain, likely due to the diverse training set including three fundamental tastes and a heterogeneous 'other' class capturing substantial chemical space. Feature analysis indicates that properties tied to charge distribution, electronegativity, and polarizability are central to taste recognition, aligning with receptor-binding mechanisms. Comparative evaluations show that VirtuousMultiTaste matches or exceeds specialized tools in their domains while offering the advantage of multi-taste capability. External screenings reveal plausible taste distributions (e.g., predominance of bitter compounds in natural products), and food composition analyses (coffee, chocolate) align with expected sensory profiles, illustrating practical applicability for screening and hypothesis generation. However, the tool predicts tastes of individual compounds rather than holistic food taste, which depends on concentrations and multisensory factors. Overall, the findings demonstrate that a parsimonious, explainable multi-class model can effectively predict multiple tastes and support data-driven food design and discovery.

VirtuousMultiTaste is a multi-class machine learning predictor that identifies bitter, sweet, and umami tastes versus other tastes from molecular structure, leveraging a hybrid heuristic optimization and Random Forest classifier with a compact, interpretable 15-feature set. It delivers strong cross-validated and external test performance, outperforms or matches specialized predictors, generalizes across chemical similarity ranges, and is deployed via a user-friendly web interface with code and data availability. Future work includes expanding to all five basic tastes by incorporating sour and salty classes, enhancing model explainability with simpler descriptors or mappings to structural motifs, and developing holistic models that integrate concentrations and multisensory factors to predict overall food sensory profiles, enabling applications in nutrition, precision medicine, and food product engineering.

• Limited number of experimentally confirmed umami compounds constrained training and external evaluation (no umami in the external test set), potentially affecting generalizability for umami.
• Sour and salty classes were excluded due to data scarcity, limiting coverage of the full taste spectrum.
• The ‘Other’ class aggregates diverse tastes (including tasteless), introducing heterogeneity that may blur class boundaries.
• Reliance on 2D descriptors may miss information present in 3D conformations; although 2D performed well, some interactions are 3D-dependent.
• The model predicts the taste of isolated compounds, not overall food taste, which depends on concentrations, matrix effects, processing, and multisensory inputs.
• Some per-class performance metrics on the test set could not be fully assessed for umami due to the absence of umami samples in the held-out test set.