ChargeNet: E(3) Equivariant Graph Attention Network for Atomic Charge Prediction

Area of Science:

• Quantum chemistry
• Computational chemistry
• Artificial intelligence in chemistry

Background:

• Accurate atomic charge prediction is crucial for drug design and discovery.
• Quantum mechanics methods are accurate but computationally expensive for large molecules.
• Existing machine learning models for atomic charges often lack accuracy and generalization.

Purpose of the Study:

• To develop a highly accurate and generalizable machine learning model for atomic charge prediction.
• To address the limitations of current methods in terms of accuracy and computational cost.
• To enhance the efficiency of drug design and discovery pipelines.

Main Methods:

• An advanced equivariant graph attention neural network was developed.
• A global graph attention mechanism was employed to model long-range electrostatic interactions.
• Structural symmetry-preserving transformations and multiscale attention were utilized.

Main Results:

• The proposed model demonstrated over 40% average improvement in charge prediction accuracy compared to baseline models.
• A 54.6% improvement in accuracy was achieved on external RESP test datasets.
• The model outperformed OPLS3 charges and baseline deep learning models in virtual screening settings.

Conclusions:

• The developed equivariant graph attention network offers a significant advancement in atomic charge prediction.
• The model exhibits high accuracy, generalization, and robustness for complex molecular systems.
• This approach holds substantial potential for accelerating scientific discovery in areas like drug design.