End-to-End Modeling of Reaction Field Energy Using Data-Driven Geometric Graph Neural Networks.

Area of Science:

• Computational chemistry and molecular modeling.
• Biophysics and structural biology.
• Machine learning for scientific applications.

Background:

• Electrostatic interactions are crucial for biomolecular structure, dynamics, and function.
• The Poisson-Boltzmann (PB) equation accurately models these interactions but is computationally intensive.
• Existing approximations like the Generalized Born (GB) model sacrifice accuracy for efficiency.

Purpose of the Study:

• To develop a computationally efficient and accurate method for calculating PB electrostatic energies.
• To overcome the limitations of traditional PB solvers and GB approximations.
• To enable precise electrostatic modeling for large biomolecules and small molecules in drug discovery.

Main Methods:

• Developed PBGNN, a novel end-to-end framework using geometric graph neural networks.
• Incorporated sinusoidal embeddings of atomic charges and a message-passing architecture.
• Introduced a charge-weighted mean squared error (CMSE) objective to stabilize training.

Main Results:

• PBGNN achieves high accuracy in predicting PB energy with linear computational complexity.
• Demonstrated reliable and precise PB free energy predictions for biomolecular complexes and small molecules.
• Showcased strong generalizability, scalability, and potential for drug discovery applications.

Conclusions:

• PBGNN offers a scalable and accurate alternative for electrostatic modeling, surpassing GB approximations.
• The framework's performance on diverse datasets highlights its utility in computational chemistry and drug discovery.
• Open-source release of PBGNN facilitates further research in accurate electrostatic analysis.