DeepAtomicCharge: a new graph convolutional network-based architecture for accurate prediction of atomic charges

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Atomic charges are crucial for drug-target interactions.
• High-level quantum mechanics (QM) calculations for atomic charges are computationally expensive.
• Existing machine learning (ML) methods often rely on predefined molecular properties, leading to potential biases.

Purpose of the Study:

• To develop a novel, high-accuracy, and efficient atomic charge prediction model.
• To overcome the limitations of knowledge-dependent feature selection in ML-based charge prediction.
• To enable large-scale structure-based drug design through improved atomic charge calculations.

Main Methods:

• Developed DeepAtomicCharge, a model utilizing a graph convolutional network (GCN) architecture.
• The GCN dynamically learns molecular features from atomic properties and connectivity, requiring no prior knowledge.
• Implemented a novel GCN architecture designed for efficient feature extraction and prediction.

Main Results:

• Achieved an average root-mean-square error (RMSE) of 0.0121 e, significantly outperforming previous benchmarks (0.0180 e).
• Demonstrated superior prediction accuracy on external test sets.
• The GCN architecture requires substantially less storage space compared to other methods.

Conclusions:

• DeepAtomicCharge offers a highly accurate and efficient method for predicting atomic charges.
• The model's ability to dynamically learn features eliminates knowledge-dependent biases.
• This approach facilitates large-scale structure-based drug design by providing fast and reliable atomic charge predictions.