Multilevel Fusion Graph Neural Network for Molecule Property Prediction

Area of Science:

• Computational chemistry
• Machine learning for drug discovery
• Molecular representation learning

Background:

• Accurate prediction of molecular properties is crucial for drug discovery.
• Existing graph neural networks (GNNs) face challenges in capturing both local and global molecular structures simultaneously.
• There is a need for advanced models to improve molecular representation learning.

Purpose of the Study:

• To propose a novel Multilevel Fusion Graph Neural Network (MLFGNN) for improved molecular property prediction.
• To jointly model local and global molecular dependencies using integrated Graph Attention Networks and a Graph Transformer.
• To enhance molecular representation by incorporating molecular fingerprints and adaptive fusion mechanisms.

Main Methods:

• Developed a Multilevel Fusion Graph Neural Network (MLFGNN).
• Integrated Graph Attention Networks and a novel Graph Transformer for joint local and global dependency modeling.
• Incorporated molecular fingerprints as a complementary modality with an attention interaction mechanism for adaptive fusion.

Main Results:

• The MLFGNN consistently outperformed state-of-the-art methods on multiple benchmark datasets for both classification and regression tasks.
• Extensive experiments validated the model's superior performance in molecular property prediction.
• Interpretability analysis confirmed the model's ability to capture task-relevant chemical patterns.

Conclusions:

• The proposed MLFGNN demonstrates significant improvements in molecular property prediction accuracy.
• Multilevel and multimodal fusion strategies are effective for enhancing molecular representation learning.
• The model's interpretability supports its utility in understanding chemical patterns relevant to specific tasks.