Performance of Universal Machine-Learned Potentials with Explicit Long-Range Interactions in Biomolecular Simulations

Area of Science:

• Computational chemistry
• Machine learning in materials science
• Biophysics

Background:

• Universal machine-learned potentials (MLPs) offer transferable accuracy across various molecular properties.
• Their application to complex biomolecular simulations is still an emerging research area.

Purpose of the Study:

• To systematically evaluate equivariant message-passing neural networks for biomolecular simulations.
• To assess the impact of model size, training data, and electrostatic treatments on simulation accuracy.

Main Methods:

• Trained equivariant message-passing architectures on the SPICE-v2 dataset.
• Evaluated models on benchmark datasets and simulations of water, NaCl solutions, and biomolecules (alanine tripeptide, Trp-cage, Crambin).
• Assessed the influence of model size, training data composition, and long-range electrostatic treatments.

Main Results:

• Larger models improved benchmark accuracy but not consistently simulation properties.
• Training data composition significantly influenced predicted properties.
• Long-range electrostatics had no systematic impact, though increased conformational variability for Trp-cage.

Conclusions:

• Current universal MLPs face challenges in biomolecular simulations due to imbalanced datasets and evaluation practices.
• Further development is needed to ensure reliable and transferable accuracy for complex biological systems.