Machine Learning Screen Finds Everyday Flavor Compounds That Activate Human Irritation Receptor TRPA1

Everyday flavor compounds such as vanillin, eugenol, and thymol can activate the human TRPA1 ion channel, the molecular sensor behind the tingling, burning, and stinging sensations known as chemesthesis, according to a new study published in Current Research in Food Science. The finding comes not from a chance discovery but from a deliberately engineered pipeline that combines machine learning prediction with laboratory validation, offering a template for how computational screening could reshape the search for flavor-active molecules.

Chemesthesis is the often-overlooked third pillar of flavor perception, alongside taste and smell. Mediated by the trigeminal nervous system, it converts specific chemical components in food into physical sensations such as tingling, burning, cooling, or numbness. These sensations lend intensity, dynamics, and spatial complexity to flavor, and they also contribute to eating motivation and food acceptability. The sensation relies heavily on the transient receptor potential (TRP) channel family, a group of non-selective cation channels expressed in trigeminal nerve fibers and epithelial cells. Among these subtypes, TRPA1 stands out for its extraordinary sensitivity to a wide range of irritant compounds, making it a key molecular hub in chemosensory perception.

Yet that same promiscuity has long frustrated researchers trying to discover new TRPA1 ligands. Unlike many receptors that follow a lock-and-key binding model, TRPA1 can be activated by structurally diverse compounds, partly because its N-terminus carries multiple electrophilic cysteine and lysine residues that many agonists modify covalently. Non-electrophilic agonists can also activate the channel through non-covalent binding that induces calcium influx. Traditional molecular docking algorithms, which typically assume fixed binding pockets and specific hydrogen bonds or hydrophobic interactions, cope poorly with these covalent and non-specific activation mechanisms. Human sensory testing adds further complications: because TRPA1 activation is often accompanied by pain or stinging, ethical constraints limit high-throughput testing at high concentrations, and significant individual differences in chemosensory sensitivity make it difficult to establish standard ligand thresholds.

To overcome these obstacles, the research team, led by Xiaoming Wen and colleagues, built a screening pipeline that integrates computational prediction with in vitro experimental validation. The dataset was drawn from the public bioactivity database ChEMBL, with small molecules explicitly annotated as TRPA1 agonists forming the positive set and molecules identified as inactive through high-throughput screening, along with background controls, forming the negative set. After standardization of the molecular structures using the RDKit toolkit, including desalting, protonation state unification, and tautomer standardization, each compound was converted into a 2048-bit Morgan topological fingerprint encoding the presence or absence of specific chemical substructures.

The team compared five classical machine learning algorithms: Random Forest (RF), Support Vector Machine, Gradient Boosting Classifier, Logistic Regression, and k-Nearest Neighbors. Using stratified five-fold cross-validation with a multi-metric evaluation framework, the Random Forest model emerged as the most balanced performer, achieving an area under the ROC curve of 0.94 plus or minus 0.03 and maintaining high levels across accuracy, F1-score, and recall with relatively stable 95 percent confidence intervals. Because the primary goal was to prioritize potential TRPA1 agonists rather than simply classify compounds, the model’s ability to minimize the risk of missing active compounds weighed heavily in its favor. Early enrichment analysis further supported this choice: at low database screening ratios, the RF model enriched a high proportion of true agonists, with an enrichment factor at 5 percent of approximately 1.9, significantly better than random screening.

Chemical space analysis revealed why the problem is so difficult and why machine learning can help. Agonists and inactive compounds showed partial overlap in overall chemical space but also several relatively enriched regions, and the two classes differed significantly in molecular weight, lipophilicity, and topological polar surface area. Agonists tended toward lower molecular weight distributions, suggesting smaller molecules may more easily enter or adapt to the ligand-binding region of TRPA1. The skeletal distribution of active compounds was strikingly diverse: acyclic skeletons accounted for roughly 24.76 percent of the agonist set, followed by benzene derivatives at 14.29 percent and dibenzazepine derivatives at 9.05 percent. This structural diversity means model predictions may still be influenced by the known distribution of active scaffolds, a limitation the authors acknowledged when selecting candidates for experimental confirmation.

Based on prediction scores, applicability domain confidence, scaffold distribution, and experimental feasibility, the team selected six representative flavor compounds for in vitro testing: vanillin, eugenol, thymol, acetic acid, lactic acid, and citric acid. The three aromatic compounds showed high prediction scores and applicability-domain confidence, while predictions for the acidic compounds varied more. To validate the predictions, the researchers constructed a HEK293T cell model stably expressing human TRPA1, introduced via recombinant lentivirus and confirmed through Western blotting, which detected a TRPA1 protein band of approximately 130 kilodaltons, along with quantitative PCR and immunofluorescence staining showing co-localized GFP and TRPA1 signals.

Calcium flux assays using FLIPR Calcium 6 dye confirmed that the machine learning predictions held up in living cells. Vanillin activated hTRPA1 with a half-maximal effective concentration of 24.11 micromolar, eugenol with an EC50 of 22.42 micromolar and a higher maximum response intensity than vanillin, and thymol with an EC50 of 49.02 micromolar. All three showed clear dose-dependent activation in hTRPA1-expressing cells, while untransduced HEK293T cells showed weaker or near-background responses to the same compounds, indicating the responses were primarily dependent on TRPA1 expression. The three organic acids behaved differently, showing pH-dependent activation windows between pH 7.4 and 5.0. Acetic acid and lactic acid produced specific activation signals distinguishable from wild-type cells in the pH range of roughly 6.5 to 6.0, while citric acid, a polycarboxylic acid with three carboxyl groups, showed a narrower specificity signal-to-noise ratio, possibly because it penetrates the hydrophobic lipid bilayer more slowly than monocarboxylic acids in its undissociated state, resulting in slower intracellular acidification and a weakened peak in TRPA1-mediated calcium influx.

Molecular docking using AutoDock Vina against the human TRPA1 receptor structure resolved by cryo-electron microscopy revealed that the aromatic compounds and organic acids occupy different local binding regions and stabilize the complex through different interaction modes. Vanillin interacts with residues including Leu609, Phe612, Cys621, and Lys661, forming pi-alkyl or alkyl interactions with hydrophobic residues while its oxygen-containing functional groups participate in polar interactions. Eugenol binds near residues including Ala718, His719, and Arg852, with its phenolic hydroxyl group and aromatic ring providing the structural basis for hydrogen bonding and hydrophobic interactions respectively. Thymol interacts with residues including Phe884, Leu936, and Phe947, its hydrophobic substituents enhancing binding stability in the hydrophobic pocket. The acidic compounds rely more heavily on polar interactions and hydrogen bond networks, with citric acid forming multi-point hydrogen bonds or electrostatic interactions with residues including Lys787, Gln791, and Arg852.

Molecular dynamics simulations over 100 nanoseconds in a POPC lipid bilayer at 310 Kelvin supported the dynamic stability of these complexes. Root mean square deviation values stabilized with a main fluctuation range of approximately 0.4 to 0.6 nanometers, and radii of gyration remained within approximately 4.55 to 4.70 nanometers, indicating no significant change in overall protein compactness. Hydrogen bond analysis showed that acidic compounds, especially lactic acid and citric acid, formed more transient hydrogen bonds during the simulation, while the aromatic ligands formed fewer, suggesting their stable binding relies more on hydrophobic, aromatic, and van der Waals interactions. These results were largely consistent with the docking findings. The authors note that the pipeline establishes a foundation for future high-throughput approaches to targeted flavor design and sensory receptor research, though the findings so far rest on cell-based assays rather than human sensory data, and the model’s predictions remain influenced by the distribution of known active scaffolds in the training data.

Subject of Research: Machine learning prediction and in vitro validation of TRPA1-activating flavor compounds

Article Title: Integrated Machine Learning and In Vitro Pipeline for TRPA1-Active Chemesthetic Flavor Compounds: From Predictive Screening to Molecular Mechanisms

Article References: Wen, X., Zhang, Q., Wang, S., Shi, Q., Wang, J., Fan, W., Guan, S., & Chai, G. (2026). Integrated machine learning and in vitro pipeline for TRPA1-active chemesthetic flavor compounds: From predictive screening to molecular mechanisms. Current Research in Food Science, 13, Article 101611. https://doi.org/10.1016/j.crfs.2026.101611

Image Credits: AI Generated

DOI: 10.1016/j.crfs.2026.101611

Keywords: TRPA1, chemesthesis, machine learning, random forest, flavor compounds, vanillin, eugenol, thymol, calcium flux assay, molecular docking, molecular dynamics, HEK293T cells