Coordinate-Free and Low-Order Scaling Machine Learning Model for Atomic Partial Charge Prediction for Any Size of...

Area of Science:

• Computational chemistry
• Machine learning applications in molecular modeling

Background:

• Atomic partial charge is vital in chemistry and drug-target recognition.
• Traditional quantum methods for calculating atomic charges are computationally expensive and slow.
• Existing machine learning models face limitations such as requiring high-accuracy geometry optimization or restricting molecule size.

Purpose of the Study:

• To develop a novel machine learning model for rapid and accurate prediction of atomic partial charges.
• To overcome the computational cost and size limitations of existing atomic charge calculation methods.

Main Methods:

• A message-passing featurizer was developed to extract atomic environment information based on molecular connectivity.
• This featurizer was integrated with a neural network for efficient prediction of atomic partial charges.
• The model's performance was evaluated using Hirshfeld charge prediction.

Main Results:

• The proposed model achieves a root-mean-square error of 0.018e in Hirshfeld charge prediction.
• The model demonstrates efficiency with an overall time complexity of O(n^2).
• The model automatically adapts to molecules of varying sizes without preprocessing limitations.

Conclusions:

• The developed machine learning model offers a significant speed-up for atomic charge calculations.
• This approach broadens the applicability of atomic charge analysis in diverse scientific fields and scenarios.
• The model provides an efficient and scalable solution for predicting atomic partial charges.