Developing a Differentiable Long-Range Force Field for Proteins with E(3) Neural Network-Predicted Asymptotic...

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning

Background:

• Accurately describing long-range interactions is crucial for molecular dynamics (MD) simulations of proteins.
• High-quality long-range potentials are essential for range-separated machine learning force fields.

Purpose of the Study:

• To develop a comprehensive database of asymptotic parameters for protein fragments.
• To create an accurate and transferable E(3) neural network (E3NN) model for predicting these parameters.
• To enable next-generation protein force fields capable of describing long-range interactions.

Main Methods:

• Construction of an asymptotic parameter database using active learning, covering atomic multipole moments, polarizabilities, and dispersion coefficients.
• Development of an E(3) neural network (E3NN) to predict asymptotic parameters directly from local protein fragment geometry.
• Validation against symmetry-adapted perturbation theory (SAPT) for electrostatic and dispersion energies.

Main Results:

• The database effectively represents protein fragments' conformational diversity with 78,000 data points.
• E3NN models achieved an R-squared value of 0.999 for predicting asymptotic parameters across protein fragments and dipeptide test sets.
• Predicted long-range electrostatic and dispersion energies showed minimal error (0.07 and 0.02 kcal/mol, respectively) compared to SAPT.

Conclusions:

• The developed force fields accurately describe long-range interactions in proteins.
• The E3NN models demonstrate high accuracy and transferability for asymptotic parameter prediction.
• This work paves the way for more accurate and efficient protein simulations using next-generation force fields.