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

Anna Katharina Picha, Sara Tkaczyk, Thierry Langer, Marcus Wieder, Stefan BoreschThe Journal of Physical Chemistry Letters 2025 · OpenReview

TL;DR

This paper introduces a novel, architecture-independent method for calculating absolute solvation free energies using neural network potentials (NNPs). By manipulating the NNP's neighbor list, the authors enable the "decoupling" of atoms and molecules, a crucial step in alchemical free energy simulations, without needing to modify the NNP's internal architecture. This opens the door for more accurate and efficient free energy calculations with the growing class of NNPs.

What Problem Does This Paper Address?

Calculating free energy differences (ΔG) is fundamental to understanding chemical processes, from reaction feasibility to drug binding. Alchemical free energy simulations (FES) are a powerful computational tool for this, but they often rely on molecular mechanics (MM) force fields, which can limit accuracy. While quantum mechanics (QM) offers higher accuracy, it's computationally too expensive for large-scale FES. Neural network potentials (NNPs) have emerged as a promising middle ground, offering near-QM accuracy at a fraction of the computational cost.

However, a significant challenge in applying NNPs to FES, particularly for absolute solvation free energies (ASFE), is the "decoupling" or "annihilation" of atoms or molecules. This process involves gradually turning off interactions, which is straightforward with traditional force fields using "soft-core" potentials. NNPs, however, are trained on specific physical interactions and can struggle with these nonphysical intermediate states. Existing methods to address this often require architecture-specific modifications to the NNP, limiting their general applicability. This paper tackles the problem of performing alchemical transformations with NNPs in an architecture-independent manner.

Methods

The core innovation of this paper lies in its "4D-Decoupling" approach, which leverages the universal requirement of NNPs to construct a neighbor list. This list identifies atom pairs within a certain cutoff distance (rcut) that the NNP will consider for energy and force calculations.

The authors propose modifying the interatomic distances (dij) between solute and solvent atoms as a function of a coupling parameter (λ), which ranges from 0 (fully coupled) to 1 (fully decoupled). Specifically, they introduce a shifting function, f(dij, rcut, λ), that effectively increases the distance between solute and solvent atoms as λ increases. At λ=1, this modified distance becomes greater than rcut, ensuring that these solute-solvent pairs are excluded from the NNP's neighbor list and thus their interactions are effectively turned off.

This "4D-Decoupling" is architecture-independent because it operates on the neighbor list, a fundamental component of any NNP, rather than on the NNP's internal structure. The paper explores four different shifting schemes, two linear and two "4D" (inspired by previous work), to achieve this decoupling. They also employ a validation strategy based on "cycle closure," where the free energy difference between two states is calculated via two independent pathways (absolute solvation and relative transformation) and the results must agree.

Datasets

The paper does not introduce new datasets for training NNPs. Instead, it utilizes pre-trained MACE-OFF23(S) and MACE-OFF23(M) models.

For evaluation, they use:

  • Small molecule solvation free energies: They report ASFE for water, methane, ethane, methanol, ethanol, toluene, and phenol. Experimental values are taken from the FreeSolv database and the Minnesota solvation database.
  • Tautomer pairs: Two specific tautomer pairs (tp1516 and tp558) are used to test the cycle closure validation strategy. The relative free energy calculations for these pairs are based on a previously established methodology.

Key Insights and Ideas

  • Neighbor list as a universal hook: The paper highlights that the neighbor list construction is a common denominator across all NNP architectures, making it an ideal point of intervention for alchemical transformations.
  • Architecture-independent decoupling: The proposed method avoids the need for NNP-specific modifications, significantly broadening the applicability of FES to a wider range of NNPs.
  • "4D-Decoupling" mimics soft-core potentials: By manipulating distances, the approach effectively creates a "soft-core" behavior for NNPs, allowing for smooth transitions between physical and nonphysical states.
  • Cycle closure as a robust validation: The use of cycle closure for validation provides a self-consistent check of the method's correctness without relying on experimental data, which can be prone to its own uncertainties.

Key Results

  • The authors successfully demonstrate the "4D-Decoupling" method for calculating absolute solvation free energies using the MACE-OFF23(S) and MACE-OFF23(M) NNPs.
  • Cycle closure errors for tautomer pairs were found to be small (0.0 ± 0.4 kcal/mol and -0.2 ± 0.2 kcal/mol), indicating the correctness of the method.
  • Solvation free energies for small molecules calculated with MACE-OFF23(S) showed good agreement with experimental values, with the largest deviation for phenol (0.6 kcal/mol).
  • Interestingly, the smaller MACE-OFF23(S) model generally performed better than the MACE-OFF23(M) model for ASFE calculations, suggesting that model size doesn't always correlate with improved condensed-phase accuracy.

Limitations and Future Work

  • NNP limitations outside training domain: The authors acknowledge that current NNPs may be unreliable when used outside their training domain (e.g., for properties not explicitly trained on). This means that agreement with experiment for ASFE should be interpreted cautiously.
  • Simulation instabilities: Occasionally, simulations using NNPs can be unstable, even without alchemical modifications. The authors had to employ restraints to prevent issues like deprotonation in their tautomer calculations.
  • Optimization of λ-scheduling: The optimization of the λ-scheduling (how the coupling parameter is advanced) was only cursorily attempted, suggesting that further work in this area could improve efficiency and accuracy.
  • Combining masking and decoupling: For more complex transformations involving larger groups of atoms, a combination of masking and decoupling will be necessary.
  • Generalizability to other NNPs: While the method is architecture-independent in principle, its practical implementation and performance may vary across different NNP architectures and training sets. Further validation with a broader range of NNPs is needed.

Summary updated

Abstract

Click to copy section linkSection link copied!
BibTeX
@article{picha2025architecture,
  author = {Anna Katharina Picha and Sara Tkaczyk and Thierry Langer and Marcus Wieder and Stefan Boresch},
  title = {{Architecture-Independent Absolute Solvation Free Energy Calculations with Neural Network Potentials}},
  journal = {The Journal of Physical Chemistry Letters},
  year = {2025},
  url = {https://openreview.net/forum?id=eW7yR8vXL6},
}
Figure