Generalized Born radii computation using linear models and neural networks.

Area of Science:

• Computational Chemistry
• Biophysics
• Molecular Dynamics

Background:

• Implicit solvent models are crucial for simulating biomolecular systems.
• Accurate computation of generalized Born (GB) radii is essential for these models.
• Current methods for GB radii calculation are computationally intensive and can be slow.

Purpose of the Study:

• To develop faster and more accurate methods for computing generalized Born (GB) radii.
• To leverage machine learning to predict GB radii based on molecular structure and physics.
• To improve the efficiency of classical molecular dynamics simulations.

Main Methods:

• Developed a linear model and a neural network approach for GB radii prediction.
• Input features include atomic element and neighbor atom counts within a defined radius.
• Trained models using existing accurate GB radii calculations as ground truth.

Main Results:

• The linear model is 8x faster than reference methods with a 0.94 correlation coefficient.
• Neural networks further enhance accuracy, achieving a 0.97 correlation coefficient.
• The neural network model shows a ~20% reduction in root mean square error compared to reference methods.

Conclusions:

• Machine learning models offer a significant speed and accuracy improvement for GB radii computation.
• These new methods can enhance the efficiency and applicability of molecular dynamics simulations.
• Open-source implementations are provided for both linear and neural network models.