High-Precision Atomic Charge Prediction for Protein Systems Using Fragment Molecular Orbital Calculation and Machine...

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Accurate atomic partial charges are essential for molecular simulations and drug design.
• Existing methods can be computationally expensive or lack accuracy in capturing electronic effects.
• Electronic polarization significantly impacts molecular interactions and properties.

Purpose of the Study:

• To develop highly accurate and computationally efficient neural network models for predicting atomic partial charges.
• To incorporate the effects of electronic polarization into charge prediction models.
• To validate the models on diverse and large biomolecular systems relevant to drug design.

Main Methods:

• Training neural network models using quantum mechanics data from the fragment molecular orbital (FMO) method.
• Utilizing high-quality, large-scale datasets for model training and validation.
• Developing models capable of accounting for system-dependent electronic polarization.

Main Results:

• Achieved high accuracy in predicting atomic partial charges with neural network models.
• Demonstrated low computational cost for the developed prediction models.
• Successfully predicted charges for complex protein systems, including a large biomolecule (~2000 atoms).

Conclusions:

• The developed neural network models provide accurate and efficient atomic partial charge predictions.
• The inclusion of electronic polarization enhances the applicability of these models in drug design.
• These models are expected to be widely adopted for structure-based drug design tasks.