Explainable Graph Neural Networks Reveal Molecular Features Associated with Umami Taste
- 1Faculty of Medicine, Tottori University, 86 Nishi-cho, Yonago, Japan, 683-8503
Abstract
Umami is one of the five basic tastes, and understanding the relationship between molecular structure and umami perception is important for food science. Although machine learning models have been developed to predict umami taste, the molecular features underlying their predictions remain insufficiently understood. In this study, we developed a graph neural network (GNN) model for umami taste prediction and investigated the structural features contributing to its predictions using Integrated Gradients (IG). Three datasets were constructed from ChemTastesDB, comprising 74 umami compounds and different sets of non-umami compounds. Under five-fold Murcko scaffold-disjoint cross-validation, the model achieved mean receiver operating characteristic areas under the curve of 0.935, 0.950, and 0.910 for the strict, full primary, and physicochemically matched datasets, respectively. Atom-level IG analysis revealed that functional-group prevalence did not necessarily correspond to positive attribution toward umami prediction. Notably, carboxylic acids occurred at similar frequencies in umami and non-umami compounds but showed significantly greater positive attribution in umami compounds. Furthermore, analysis of 31 structurally similar umami–nonumami molecular pairs demonstrated that regions specific to umami compounds generally received greater positive attribution than the corresponding regions in non-umami compounds. These findings suggest that the GNN captures molecular features associated with umami taste in a manner that depends on their surrounding structural context. Combining scaffold-disjoint prediction with atom-level attribution analysis provides an approach for investigating structure–taste relationships and identifying molecular features potentially relevant to umami perception.
Related
- Explainable spatiotemporal Graph Neural Networks: Multi-City PM2.5 forecasting in Bangladesh with geographic graph topology validation
- Integration of Network Toxicology and Machine Learning to Prioritize Candidate Molecular Features Associated with NNK-Related Lung Cancer
- Structural Resolution in Feature Attribution Analysis of Molecular Graph Neural Networks
- Out-of-Distribution Inverse Design of Elastic Networks with Differentiable Graph Neural Network Molecular Dynamics
- Comparative Evaluation of Graph Neural Networks for Molecular Lipophilicity Prediction
BibTeX
@misc{iwata2026explainable,
title = {{Explainable Graph Neural Networks Reveal Molecular Features Associated with Umami Taste}},
author = {Hiroaki Iwata},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010977/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010977/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Explainable Graph Neural Networks Reveal Molecular Features Associated with Umami Taste
AU - Hiroaki Iwata
PY - 2026
DA - 2026/10/23
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010977/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010977/v1
AB - Umami is one of the five basic tastes, and understanding the relationship between molecular structure and umami perception is important for food science. Although machine learning models have been developed to predict umami taste, the molecular features underlying their predictions remain insufficiently understood. In this study, we developed a graph neural network (GNN) model for umami taste prediction and investigated the structural features contributing to its predictions using Integrated Gradients (IG). Three datasets were constructed from ChemTastesDB, comprising 74 umami compounds and different sets of non-umami compounds. Under five-fold Murcko scaffold-disjoint cross-validation, the model achieved mean receiver operating characteristic areas under the curve of 0.935, 0.950, and 0.910 for the strict, full primary, and physicochemically matched datasets, respectively. Atom-level IG analysis revealed that functional-group prevalence did not necessarily correspond to positive attribution toward umami prediction. Notably, carboxylic acids occurred at similar frequencies in umami and non-umami compounds but showed significantly greater positive attribution in umami compounds. Furthermore, analysis of 31 structurally similar umami–nonumami molecular pairs demonstrated that regions specific to umami compounds generally received greater positive attribution than the corresponding regions in non-umami compounds. These findings suggest that the GNN captures molecular features associated with umami taste in a manner that depends on their surrounding structural context. Combining scaffold-disjoint prediction with atom-level attribution analysis provides an approach for investigating structure–taste relationships and identifying molecular features potentially relevant to umami perception.
ER -