Architecture-Independent Absolute Solvation Free Energy Calculations with Neural Network Potentials

Area of Science:

• Computational Chemistry
• Materials Science

Background:

• Alchemical free energy simulations (FES) are crucial for predicting molecular properties.
• Traditional FES methods rely on force fields, but neural network potentials (NNPs) offer higher accuracy.
• A key challenge in FES is the gradual decoupling of atoms and molecules.

Purpose of the Study:

• To develop a novel method for decoupling atoms and molecules in FES using NNPs.
• To ensure the method's compatibility with diverse NNP architectures.
• To validate the method's accuracy and applicability in FES.

Main Methods:

• Developed a technique to gradually decouple atoms/molecules by manipulating the neighbor list within NNP-described systems.
• Demonstrated the equivalence of this neighbor list manipulation to soft-core potentials used in force-field FES.
• Validated the method through cycle closure tests and calculation of solvation free energies.

Main Results:

• Successfully implemented a method for gradual decoupling in NNP-based FES.
• Showed that neighbor list manipulation effectively mimics soft-core potentials.
• Achieved accurate solvation free energies using the MACE-OFF23(S/M) NNP.

Conclusions:

• The proposed method provides a versatile approach for NNP-based FES.
• This technique is agnostic to specific NNP architectures, enhancing its broad applicability.
• The method enables accurate free energy calculations with potentially higher accuracy NNPs.