Transferability of gas-phase–trained NNPs to condensed-phase simulations

Determine the extent to which neural network potentials trained and primarily evaluated on gas-phase datasets are transferable to condensed-phase molecular simulations typical of computational drug discovery, assessing whether their accuracy and stability persist in condensed-phase environments.

Background

Neural network potentials are often trained and validated using gas-phase datasets and error metrics on limited sets of molecules and conformations. The paper emphasizes that many real-world applications in drug discovery require reliable behavior in the condensed phase, where multi-body interactions and long-range effects are prominent.

To address this concern, the authors perform a series of tests including gas-phase normal mode analyses, molecular dynamics of a drug-like molecule, water dimer scans, and condensed-phase water simulations. Their results show that performance can vary substantially across phases, underscoring the need to establish transferability beyond gas-phase benchmarks.

References

However, most of the model testing is performed in the gas-phase using a limited set of molecules and conformations; therefore, it is unclear how far the models are transferable to other application areas such as simulations in the condensed phase, commonplace in computational drug discovery.

— Basic stability tests of machine learning potentials for molecular simulations in computational drug discovery  (2503.11537 - Ranasinghe et al., 14 Mar 2025) in Section 1 (Introduction)