DeepChargePredictor: a web server for predicting QM-based atomic charges via state-of-the-art machine-learning...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning applications

Background:

• High-level quantum mechanics (QM) methods provide accurate atomic charges but are computationally expensive.
• This limits their application in large-scale molecular modeling, including high-throughput virtual screening.
• Machine learning (ML) approaches have emerged to approximate QM-level atomic charges efficiently.

Purpose of the Study:

• To introduce DeepChargePredictor, a web server for generating high-level QM atomic charges for small molecules.
• To integrate advanced ML algorithms (AtomPathDescriptor and DeepAtomicCharge) for charge prediction.
• To evaluate the performance of predicted charges in drug design applications.

Main Methods:

• Development of the DeepChargePredictor web server.
• Implementation of two ML algorithms: AtomPathDescriptor and DeepAtomicCharge.
• Prediction of RESP, AM1-BCC, and DDEC atomic charges.
• Comprehensive performance evaluation in end-point binding free energy calculations and virtual screening.

Main Results:

• DeepChargePredictor successfully generates high-level QM atomic charges using ML.
• The predicted charges demonstrate reliable performance in structure-based drug design.
• Performance in large-scale applications like virtual screening is comparable or superior to baseline methods.

Conclusions:

• DeepChargePredictor provides an efficient solution for obtaining accurate atomic charges in drug design.
• The web server facilitates the use of high-quality atomic charges in large-scale molecular modeling.
• This tool enhances the capabilities of virtual screening and binding free energy calculations.