Targeted free energy estimation via learned mappings.

Area of Science:

• Computational Chemistry
• Statistical Mechanics
• Machine Learning

Background:

• Free energy perturbation (FEP) is a foundational method for calculating free energy differences in molecular simulations.
• A key limitation of FEP is the need for sufficient overlap between sampling distributions, often difficult to achieve.
• Targeted FEP aims to improve overlap using high-dimensional mappings, but designing these mappings is challenging.

Purpose of the Study:

• To reframe Targeted FEP as a machine learning problem, enabling neural networks to learn optimal mappings.
• To develop a neural network architecture that handles symmetries common in atomistic simulations.
• To demonstrate significant variance reduction in free energy estimates using the proposed method.

Main Methods:

• Parameterizing the high-dimensional mapping in Targeted FEP using a neural network.
• Optimizing the neural network to maximize the overlap between sampling distributions.
• Developing a neural network architecture incorporating permutational and periodic symmetries.
• Testing the method on a fully periodic solvation system.

Main Results:

• The machine learning-based Targeted FEP method achieved substantial variance reduction in free energy estimates.
• The approach effectively increased the overlap between distributions.
• The method demonstrated successful application on a complex, fully periodic system.
• No additional simulation data was required beyond standard FEP inputs.

Conclusions:

• Casting Targeted FEP as a machine learning problem offers a powerful and tractable solution to the distribution overlap challenge.
• The developed neural network architecture effectively handles system symmetries, making it suitable for atomistic simulations.
• This approach significantly enhances the accuracy and efficiency of free energy calculations.