Machine Learning Modeling of Olfactory Thresholds of Pyrazine Derivatives Using Cheminformatics Descriptors

Abstract

Pyrazine derivatives represent an important class of heterocyclic compounds widely used for their potent odor properties and their numerous applications in industry and perfumery. However, predicting their olfactory thresholds remains challenging due to the complex relationship between molecular structure and sensory perception. In this study, a machine learning strategy was developed to quantitatively model the structure-odor relationships of 78 substituted 1,4-pyrazine derivatives using cheminformatics descriptors. Molecular features were calculated from SMILES representations of the molecules using RDKit and Mordred, followed by descriptor preprocessing and dimensionality reduction. Several algorithms were evaluated, including Decision Tree, Random Forest, Gradient Boosting, Bagging, Extra Trees, XGBoost, LightGBM, and Histogram-Based Gradient Boosting. In particular, Extra Trees, Random Forest, and Bagging achieved the highest external accuracies with R² values of 0.814, 0.802, and 0.784 and low RMSE values of 0.816, 0.841, and 0.878, respectively, demonstrating strong generalization ability and reduced overfitting. Plots comparing predicted and experimental results confirmed the robustness of these models. Overall, this work highlights the effectiveness of machine learning techniques for modeling olfactory properties and provides a practical computational framework for the prediction and virtual screening of odor molecules, thereby promoting more efficient development of flavors and fragrances.