The membrane dipole potential (Ψd) is a critical modulator of ion transport and protein function, making the ability to accurately predict its modifications essential for rational drug design and membrane biophysics. While molecular dynamics (MD) simulations offer a powerful alternative to challenging in vitro experiments, their predictive accuracy is often hampered by sensitivities to simulation setups and force field parameterization. In this study, we provide a systematic evaluation of how system composition, box size, water models, and small-molecule parameterization protocols influence the calculated membrane potential. Using the CHARMM36m and AMBER (Lipid21) force fields, we demonstrate that CHARMM is notably more sensitive to box composition and finite-size effects than AMBER. We further show that the ~100 mV shift induced by 4-site water models is purely systematic; therefore, computationally efficient 3-site models remain reliable for predicting relative potential changes. Finally, we compare multiple parameterization strategies for three Ψd-modifying flavonoids (baicalein, chrysin, luteolin) and show that standard CGenFF protocols fail to capture experimental trends, whereas ffTK-refinement and AMBER-based protocols (GAFF2 and Espaloma) significantly improve accuracy. Notably, the neural network-based Espaloma demonstrated surprisingly high predictive power, marking this approach as a promising, automated alternative for future studies. Our findings provide a set of practical recommendations for establishing reliable MD protocols to predict dipole potential modifications.
Malykhina et al. (Wed,) studied this question.