Machine Learning of Analytical Electron Density in Large Molecules Through Message-Passing.

Area of Science:

• Computational chemistry
• Machine learning
• Quantum chemistry

Background:

• Machine learning (ML) in computational chemistry faces challenges with accountability and task-specific tools.
• Reproducing fundamental physical properties with ML offers a unified approach to deriving diverse molecular characteristics.

Purpose of the Study:

• To develop a machine learning model capable of accurately reproducing electron density in molecules.
• To demonstrate the model's adaptability for biomolecules and its utility in calculating various chemical properties.

Main Methods:

• Implemented a message-passing neural network (MPNN) using an analytical expansion of electron density.
• The model represents electron density using isotropic and anisotropic functions.
• Methodology was adapted for large biomolecules, including proteins.

Main Results:

• Achieved high accuracy in reproducing electron density, with only a 2.5% absolute error in complex molecular cases.
• Successfully applied the methodology to describe electron density in proteins.
• Demonstrated the ability to derive atomic charges, interaction energies, and Density Functional Theory (DFT) energies.

Conclusions:

• Electron density learning represents a significant advancement in computational chemistry.
• This approach offers a promising and versatile foundation for numerous future applications in molecular modeling and property prediction.