Dual-Branch Graph Neural Network for Predicting Molecular Odors and Discovering the Relationship Between Functional...

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Predicting molecular odors is vital for synthetic chemistry and perfumery.
• Existing models face limitations in capturing complex molecular features for odor prediction.

Purpose of the Study:

• To introduce ScentGraphX, a dual-branch graph neural network model for predicting molecular odors.
• To improve the accuracy and reliability of odor prediction using advanced machine learning techniques.

Main Methods:

• Developed a dual-branch graph neural network (ScentGraphX) incorporating transfer learning and graph attention mechanisms.
• Utilized feature and subgraph encoders to capture atomic, bond, and structural molecular properties.
• Evaluated model performance on a dataset of 4967 molecules for multi-label odor descriptor prediction.

Main Results:

• ScentGraphX demonstrated superior performance in predicting multi-label odor descriptors.
• The model achieved high precision, recall, F1-score, and AUCROC compared to existing methods.
• Interpretability analysis identified key functional groups influencing odor characteristics.

Conclusions:

• ScentGraphX effectively predicts molecular odors, outperforming current models.
• The model's architecture and feature extraction capabilities are validated by ablation studies.
• This work offers significant advancements for the perfume industry and molecular design.