Anti-Symmetric Molecular Graph Learning Approach With Residual Adaptive Network Based Fuzzy Inference System for...

Area of Science:

• Computational Chemistry
• Machine Learning
• Graph Neural Networks

Background:

• Graph neural networks (GNNs) are crucial for molecular graph learning but struggle with long-range dependencies and the over-squashing problem.
• Over-squashing compresses distant node information, degrading performance in tasks like lethal dose forecasting that require understanding both local and global molecular structures.

Purpose of the Study:

• To introduce a novel anti-symmetric fuzzy-enhanced graph learning (ASFGL) model to overcome GNN limitations in molecular graph learning.
• To enhance the capture of long-range dependencies and global structural information for improved molecular property prediction.

Main Methods:

• The proposed ASFGL model integrates an anti-symmetric transformation module based on stable graph ordinary differential equations (ODEs) to ensure non-dissipative information propagation.
• A residual adaptive neuro-fuzzy inference system (ANFIS) with bell-shaped membership functions is employed for robust, interpretable, and adaptive rule-based reasoning.
• These components work together to mitigate over-squashing, capture long-range dependencies, and refine molecular representations.

Main Results:

• The ASFGL model effectively mitigates the over-squashing issue, enabling stable propagation of information and better capture of long-range dependencies.
• ASFGL yields expressive molecular embeddings by bridging local message passing and global structural awareness, proving effective for toxicity prediction.
• Evaluations on benchmark datasets show ASFGL consistently outperforms state-of-the-art GNNs in MAE/RMSE metrics, especially in deep representation learning scenarios.

Conclusions:

• The integration of anti-symmetric dynamics and fuzzy inference systems represents a significant advancement in molecular property prediction using GNNs.
• ASFGL successfully addresses foundational challenges in GNN design, offering improved performance for complex molecular tasks like lethal dose forecasting.