Fast and accurate prediction of partial charges using Atom-Path-Descriptor-based machine learning.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Partial atomic charges are crucial for molecular modeling applications like docking and simulations.
• Accurate partial charge calculation is computationally expensive using quantum mechanics, limiting high-throughput screening.

Purpose of the Study:

• To develop a novel, accurate, and efficient method for predicting partial atomic charges in small molecules.
• To introduce the Atom-Path-Descriptor (APD) as a new molecular descriptor for machine learning models.

Main Methods:

• Developed the Atom-Path-Descriptor (APD) algorithm to characterize local chemical environments based on 3D molecular structures.
• Employed ensemble machine learning algorithms, Random Forest (RF) and Extreme Gradient Boosting (XGBoost), using APDs for partial charge prediction.
• Compared APD-based models with traditional molecular fingerprints.

Main Results:

• APD-based Random Forest models outperformed traditional molecular fingerprints in predicting partial charges for all atom types.
• Extreme Gradient Boosting models trained with APDs further improved prediction accuracy, achieving an average root-mean-square error of 0.0116 e.
• The achieved accuracy is significantly better than previously reported methods, demonstrating the promise of the APD approach.

Conclusions:

• The proposed Atom-Path-Descriptor (APD) coupled with machine learning offers a highly accurate and efficient method for partial charge assignment.
• This method is suitable for large-scale molecular modeling and high-throughput virtual screening.
• The developed software framework is publicly available for use.