ANI/EFP: Modeling Long-Range Interactions in ANI Neural Network with Effective Fragment Potentials

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Machine Learning

Background:

• Deep learning neural networks (NN) offer efficient alternatives to quantum-mechanical calculations in molecular modeling.
• Current NN models struggle with accurately representing long-range interactions, limiting their application to extended molecular systems.

Purpose of the Study:

• To develop a novel deep learning approach for molecular modeling that accurately captures long-range interactions.
• To enhance the general-use neural network ANI by integrating electrostatic potentials for improved accuracy in extended systems.

Main Methods:

• Partially retraining the ANI neural network by incorporating atomic electrostatic potentials as additional input features.
• Generating electrostatic potentials using polarizable effective fragment potentials (EFP).

Main Results:

• The new ANI/EFP network achieves kcal/mol accuracy in predicting solute-solvent interaction energies on a trained dataset.
• Demonstrates potential for predicting interaction energies in novel solvent environments not present in the training data.

Conclusions:

• The proposed ANI/EFP protocol effectively addresses the limitation of long-range interactions in deep learning molecular modeling.
• This method provides a foundation for developing highly accurate and transferable neural network potentials for complex molecular systems.