Graph Neural Networks vs. Traditional QSAR: A Comprehensive Comparison for Multi-Label Molecular Odor Prediction.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Molecular odor prediction is crucial for fragrance design, food science, and safety.
• Traditional Quantitative Structure-Activity Relationship (QSAR) methods use hand-crafted descriptors.
• Graph neural networks (GNNs) offer direct learning from molecular structures.

Purpose of the Study:

• To comprehensively compare traditional QSAR and modern GNN approaches for multi-label odor prediction.
• To evaluate various machine learning algorithms and GNN architectures on the GoodScent dataset.
• To identify optimal strategies for multi-label chemical classification.

Main Methods:

• Systematic evaluation of 23 model configurations, including Random Forest, SVM, GBDT, MLP, XGBoost, LightGBM, and GNNs (GCN, GAT, NNConv).
• Utilized the GoodScent dataset with 3304 molecules and six odor types (fruity, green, sweet, floral, woody, herbal).
• Investigated three feature-processing strategies for traditional methods and three GNN architectures.

Main Results:

• GNN models significantly outperformed traditional methods in multi-label odor prediction.
• Graph Convolutional Network (GCN) achieved the highest macro F1-score (0.5193), a 24.1% relative improvement over the best traditional method (MLP: 0.4766).
• Threshold optimization was found to be essential for effective multi-label chemical classification.

Conclusions:

• GNNs are the preferred approach for molecular property prediction tasks.
• Findings provide crucial insights for handling class imbalance in chemical informatics.
• This study advances the application of machine learning in predicting molecular odor characteristics.