Attention-based graph networks for high-precision flavor prediction
Chemical Engineering Science · 2027-02 · 339 · 125423
DOI: 10.1016/j.ces.2026.125423 ↗Kurz gefasst
Fachartikel in der Zeitschrift Chemical Engineering Science, 2027-02. Verfasst von L.P. Queiroz und 4 weiteren Autor:innen. Die Arbeit nennt 32 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: JPI CH, Foundation for Science and Technology.
Abstract (Originalsprache)
The precise and scalable prediction of molecular flavor profiles is a critical challenge in flavor engineering, driven by the vast chemical space and the limitations of time-consuming traditional sensory evaluation. To overcome the reliance on limited molecular features and leverage full structural topology, this article introduces a novel methodology framework for molecular flavor classification approach based on the Graph Attention Network (GAT). The GAT model was trained on a large dataset of 11,319 molecules to predict 333 clustered flavor descriptors. The GAT framework’s innovation lies in its ability to process the molecular structure directly as a graph, where each atom is encoded with a rich feature vector. Utilizing the GATv2 architecture, the model employs a dynamic attention mechanism that assigns varying levels of importance to neighboring atoms during message passing, providing more nuanced representations of local substructures critical to molecular perception. The model achieved high predictive performance, most notably attaining exceptional precision scores for key tastes: 0.975 for sweet and 0.65 for bitter. Classification performance was strongly correlated with descriptor frequency, indicating that data volume remains a critical factor. Although constrained by dataset imbalance and the inherent complexity of modeling molecular mixtures, the framework provides a reliable and efficient method for characterizing the sensory potential of individual molecules, serving as a vital counterpart to generative design workflows and supporting targeted discovery.
Förderung
- JPI CH
- Foundation for Science and Technology
Zitieren
APA
Queiroz, L., Friborg, P., Iwakiri, I., Ribeiro, A., & Nogueira, I. (2027). Attention-based graph networks for high-precision flavor prediction. Chemical Engineering Science, 339, 125423. https://doi.org/10.1016/j.ces.2026.125423
Vancouver
Queiroz L, Friborg P, Iwakiri I, Ribeiro A, Nogueira I. Attention-based graph networks for high-precision flavor prediction. Chemical Engineering Science. 2027;339:125423. doi:10.1016/j.ces.2026.125423
BibTeX
@article{Queiroz2027,
author = {Queiroz, L.P. and Friborg, P.I. and Iwakiri, I. and Ribeiro, A.M. and Nogueira, I.B.R.},
title = {Attention-based graph networks for high-precision flavor prediction},
journal = {Chemical Engineering Science},
year = {2027},
volume = {339},
pages = {125423},
doi = {10.1016/j.ces.2026.125423}
}
EUVIMED