Adding Physicochemical Properties to 3D Models Improves Molecular Taste Classification

Computational models capable of screening how chemical structures interact with human taste modalities offer a way to accelerate early-stage ingredient evaluation and drug design without the immediate costs, ethical safety constraints, and subjective variability of human sensory testing. In a new study published in Current Research in Food Science, computational researchers Hyemi Shin and Sang-Hee Lee investigated whether combining three-dimensional structural data with global physicochemical indicators could strengthen machine learning frameworks designed to classify sweet, bitter, and tasteless molecules. Drawing curated chemical data from public repositories, the research team implemented a graph neural network architecture known as SchNet and conditioned its spatial representations on eight standardized physical properties using feature-wise linear modulation. Their primary findings demonstrate that integrating physical properties consistently enhances the discrimination of taste categories over purely geometric three-dimensional models, though conventional two-dimensional molecular fingerprints still match or exceed deep learning performance across standard chemical groupings.

Taste perception functions physiologically as a vital gatekeeper for human nutrient intake and toxin avoidance, operating primarily through specialized cellular receptors that detect modalities such as sweetness and bitterness. Caloric signaling in foods is predominantly associated with sweet molecules, prompting consumer demand for alternative sweetening agents that minimize caloric load while maintaining palatable taste profiles. Conversely, bitter tastes frequently act as an evolutionary warning mechanism against naturally occurring toxins, making unintentional bitterness a primary obstacle in both pharmaceutical formulation and functional food development. Beyond sweet and bitter sensations, tastelessness serves an essential calibration purpose in computational modeling pipelines by establishing an inactive baseline class that helps reduce false-positive predictions. However, sensory evaluation panels remain constrained by sample toxicity risks and substantial operational expenses, establishing a clear imperative for computational pipelines that can accurately prioritize prospective molecules before laboratory synthesis.

To assemble a comprehensive benchmark dataset, the authors extracted 2,944 molecules annotated with canonical simplified molecular input line entry system strings and curated taste labels from the public ChemTastesDB repository. After excluding compounds associated with alternative sensory categories such as umami, saltiness, or ambiguous multi-taste descriptors, as well as an unreadable record and three compounds that failed three-dimensional coordinate embedding, the researchers retained 2,359 distinct molecules for three-class modeling. This multiclass dataset comprised 1,181 bitter, 976 sweet, and 202 tasteless compounds, with the sweet and bitter subsets together forming a 2,157-compound cohort for binary discrimination experiments. The team subsequently linked these molecular entries to PubChem database records using valid compound identification numbers to collect eight computed physical and structural properties: molecular weight, XLogP, topological polar surface area, hydrogen-bond donor count, hydrogen-bond acceptor count, rotatable bond count, heavy-atom count, and molecular complexity.

Because 823 of the 2,359 compounds in the three-class dataset lacked valid PubChem identifiers, their property values were missing and were imputed using median values computed strictly within training folds to avoid data leakage. For spatial representations, the canonical chemical strings were converted into single three-dimensional conformers using experimental torsion knowledge distance geometry algorithms in RDKit followed by universal force field energy minimization. The investigators emphasized that these geometries reflect computational energy approximations rather than crystallographically verified physical structures. Continuous features were scaled through median and interquartile range transformations fitted exclusively on training splits. The machine learning pipeline implemented SchNet with 128 hidden dimensions, six interaction blocks, and 100 Gaussian distance basis functions within a 10-angstrom radial cutoff. Property latent vectors produced by a shallow multilayer perceptron dynamically tuned the SchNet representations via feature-wise linear modulation, and training was directed using class-balanced focal loss and AdamW optimization.

When assessed on a fixed test partition in the binary classification setting distinguishing sweet from bitter compounds, both the purely geometric three-dimensional model and the property-integrated model attained accuracy, precision, and weighted F1-scores exceeding 0.80. Incorporating the eight physicochemical properties enhanced classification performance across all measured metrics compared with relying solely on three-dimensional atomic coordinates and interatomic distances. In an error analysis of the property-integrated model, the system successfully recognized 87.18 percent of all sweet test compounds, exhibiting a moderate asymmetric error profile in which 24.15 percent of bitter molecules were incorrectly predicted as sweet, whereas only 12.82 percent of sweet molecules were misclassified as bitter. Permutation importance tests revealed that the sweet class predictions were particularly sensitive to perturbations in hydrogen-bond donors, hydrogen-bond acceptors, and lipophilicity indices, showing weighted F1 decreases between 0.25 and 0.30.

Expanding the evaluation to the ternary classification system by introducing the 202 tasteless compounds highlighted the pronounced difficulty of modeling taste inactivity. In the fixed-split three-class test, the baseline three-dimensional model without property inputs produced an accuracy of 0.650 and a weighted F1-score of 0.656, whereas the property-conditioned architecture improved those metrics to 0.731 and 0.743, respectively. However, because the weighted F1 metric is predominantly driven by the abundant sweet and bitter classes, the researchers calculated an unweighted macro-averaged F1-score of 0.636, revealing a substantial performance gap. This disparity was driven by the minority tasteless category, where test recall reached only 46.34 percent, with 31.71 percent of tasteless compounds misclassified as sweet and 21.95 percent misclassified as bitter. Permutation analysis indicated that while hydrogen-bond acceptors, hydrogen-bond donors, and molecular complexity contributed to tasteless predictions, global properties exerted less relative influence on tastelessness than on sweetness.

To investigate whether the weak sensitivity toward tasteless molecules stemmed from severe sample sparsity in the training set, which contained only 119 original tasteless compounds, the authors applied a targeted three-dimensional data augmentation protocol. By generating up to three additional computational conformers per tasteless molecule via distance geometry and injecting minor Gaussian coordinate noise with a standard deviation of 0.02 angstroms, the team added 308 augmented structural samples without altering molecular identities or physical property values. Across repeated training iterations over five random initialization seeds using matched class-balanced mini-batch sampling, augmentation produced a modest, directionally consistent improvement in the property-integrated model. The average macro-F1 increased from 0.624 to 0.642, tasteless precision improved from 0.277 to 0.312, and tasteless F1 rose from 0.340 to 0.377. Conversely, tasteless recall changed marginally from 0.444 to 0.478, demonstrating that geometric coordinate perturbation alone could not provide substantial chemical diversity.

To contextualize their deep learning framework within broader cheminformatics methodology, the researchers benchmarked their architecture against a classical random forest baseline constructed with 2,048-bit Morgan circular fingerprints and 500 trees under an identical five-fold out-of-fold cross-validation scheme. In this deterministic out-of-fold evaluation, the conventional fingerprint random forest baseline outperformed an extended SchNet model incorporating fingerprints and properties across both tasks. For binary classification, the random forest achieved an accuracy of 0.904 and a macro-F1 of 0.903, compared with 0.875 and 0.874 for the extended SchNet model. In the three-class task, the random forest reached an accuracy of 0.841 and a macro-F1 of 0.738, whereas the extended SchNet model attained 0.788 and 0.698. A soft-vote ensemble blending the two models yielded marginal adjustments that did not achieve statistical significance over the standalone random forest under standard molecule-grouped partitioning.

A critical ablation study of seven distinct algorithmic configurations on the five-fold cross-validation splits yielded several vital mechanistic insights. Adding physicochemical properties to SchNet substantially improved its performance, lifting ternary macro-F1 from 0.588 in the spatial-only network to 0.636 in the property-conditioned network. However, feature-wise linear modulation did not show a statistically significant difference compared to simple numerical feature concatenation, yielding a minor paired bootstrap difference of plus 0.008. Furthermore, when the investigators stratified predictions by property availability, they found that property conditioning improved macro-F1 by 0.057 among the 1,536 molecules with verified PubChem values, but actually reduced macro-F1 by 0.147 among the 823 compounds where all properties had been imputed, demonstrating that the predictive gain was strictly dependent on real experimental or computed data rather than numerical median imputation.

To explore model generalizability beyond familiar molecular skeletons, the authors executed a stringent validation using Bemis-Murcko scaffold-grouped splits, ensuring that chemical scaffolds in the test sets were completely absent from the training sets. Under this rigorous evaluation, the performance of the fingerprint random forest deteriorated markedly, with its ternary macro-F1 dropping from 0.738 to 0.678 and its tasteless recall plummeting from 0.416 to 0.272. In contrast, the spatial and property-conditioned models exhibited greater resilience to unfamiliar chemical architectures. Across five training seeds on the scaffold-grouped partitions, an ensemble combining the random forest and the extended SchNet model achieved a superior ternary macro-F1 of 0.699 and elevated tasteless recall to 0.361, outperforming the standalone fingerprint baseline across every evaluated random seed and suggesting that three-dimensional structural representations provide complementary value when generalising to novel scaffolds.

Substructure perturbation analyses using SMARTS pattern masking provided chemical interpretations of how the trained neural network evaluated specific molecular motifs. Displacing the spatial coordinates of atoms within matched functional groups revealed that bitter predictions were most sensitive to sulfonamide and carboxylate motifs, each yielding an average prediction probability drop of 0.12. Sweet predictions experienced their steepest probability decreases upon the perturbation of halogen-bearing substituents (0.30 drop) and primary or secondary amine motifs (0.29 drop). For the tasteless class, masking alcohol groups and hydroxyl acidic motifs produced probability decreases of 0.24 and 0.23, respectively. The investigators noted that while these computational perturbations highlight motifs that strongly sway model decisions, they identify associative algorithmic signals rather than proving direct physiological receptor binding mechanisms or biological activation pathways.

The study highlights several key boundaries that define current computational taste prediction and outline necessary directions for future investigation. The total volume of curated taste data remains modest, and no external experimental validation dataset was available to test the model outside public repository distributions. Additionally, the computationally generated single conformers do not capture physiological conformational ensembles or solvation phenomena, and the high mutual correlation among descriptors such as molecular weight, heavy atom count, and surface area precludes isolating the influence of single physical parameters. The authors concluded that while adding physical properties to spatial graph neural networks offers measurable benefits, particularly when handling unencountered scaffolds, predicting tastelessness remains challenging due to broad structural heterogeneity, requiring richer biological assay data and more advanced representation techniques.

Subject of Research: Computational molecular taste prediction using 3D graph neural networks and physicochemical properties

Article Title: Integrating physicochemical properties into an augmented 3D molecular learning framework for taste prediction

Article References: Shin, H., & Lee, S.-H. (2026). Integrating physicochemical properties into an augmented 3D molecular learning framework for taste prediction. Current Research in Food Science, Article 101623. https://doi.org/10.1016/j.crfs.2026.101623

Image Credits: AI Generated

DOI: 10.1016/j.crfs.2026.101623

Keywords: taste prediction, computational chemistry, graph neural networks, SchNet, physicochemical properties, ChemTastesDB, PubChem, feature-wise linear modulation, sweet taste, bitter taste, tasteless compounds, conformer generation, random forest baseline, Bemis-Murcko scaffold, permutation importance, substructure perturbation, machine learning in food science