TorchANI-Amber: Bridging Neural Network Potentials and Classical Biomolecular Simulations.

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning in science

Background:

• Molecular dynamics (MD) simulations are crucial for understanding biomolecular systems.
• Traditional force fields have limitations in accuracy and transferability.
• Machine learning potentials, like ANI, offer a promising alternative for accurate energy prediction.

Purpose of the Study:

• To introduce TorchANI-Amber, an interface for integrating ANI machine learning potentials into the Amber MD simulation suite.
• To enable routine biomolecular simulations using neural network potentials with high accuracy.
• To demonstrate the extensibility and performance of the interface for various biomolecular systems.

Main Methods:

• Integration of ANI neural network potentials within the Amber software suite (sander and pmemd engines).
• Implementation of optimized CUDA routines for efficient feature vector computation.
• Extension of the interface to support other energy predicting potentials (AIMNet2, Nutmeg).
• Conducting MD simulations on biomolecular systems (ubiquitin, Trp-cage) in explicit solvent.

Main Results:

• TorchANI-Amber successfully integrates ANI potentials into Amber, supporting all Amber capabilities.
• Simulations demonstrated good energy conservation and stability for biomolecular systems.
• The interface enables large-scale simulations (hundreds of thousands of atoms) at near DFT accuracy.
• Successful application in enhanced sampling techniques like replica-exchange MD.

Conclusions:

• TorchANI-Amber provides a versatile and efficient platform for biomolecular MD simulations using machine learning potentials.
• The interface facilitates the use of high-accuracy neural network potentials, approaching DFT accuracy, in large-scale simulations.
• This work advances the application of machine learning in computational biophysics and drug discovery.