Efficient Learning of Molecular Properties Using Graph Neural Networks Enhanced with Chemistry Knowledge

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Graph neural networks (GNNs) are effective for predicting molecular properties from structural data.
• GNNs face limitations in capturing global molecular properties due to oversmoothing and expressivity challenges.
• Existing GNN models struggle to inherently learn complex chemical knowledge.

Purpose of the Study:

• To develop a GNN-based model that integrates explicit chemistry knowledge to enhance molecular property prediction.
• To investigate the impact of providing global graph information to GNNs for chemical applications.
• To compare the performance of the enhanced GNN model against pure GNNs and large foundation models.

Main Methods:

• A simple GNN architecture was designed, incorporating domain-specific chemistry knowledge.
• The model was trained on small-molecule datasets for regression tasks.
• Node-level prediction capabilities were implemented to identify important molecular substructures using SMILES encoding.
• Performance was evaluated on several benchmarks against state-of-the-art models.

Main Results:

• The GNN model integrating chemistry knowledge significantly outperformed pure GNN approaches in accuracy.
• The enhanced model demonstrated competitive or superior performance compared to larger, state-of-the-art models, including foundation models.
• The node-level prediction enabled the identification of key molecular substructures influencing predictions.
• The model achieved efficient training with modest computational resources.

Conclusions:

• Integrating chemistry knowledge into GNNs is crucial for overcoming limitations in predicting molecular properties.
• Providing GNNs with easy access to global graph information enhances their applicability in chemistry.
• The developed model offers a practical and accurate solution for molecular property prediction, suitable for widespread use.