AMOEBA-Ions: improved description of high field polarization for interactions of proteins with monovalent cations

The reliability of molecular mechanics simulations to predict ion binding to proteins depends on their ability to describe accurately both ion-protein and ion-water interactions. Ion parameters are typically constructed using reference data on ion-water interactions. Many studies show that this strategy produces large errors in ion-protein interactions. Studies also show that this transferability error can be corrected to some extent by assigning separate sets of non-bonded cross-terms for every distinct pair of ion type and its coordinating protein functional group. Here we use the polarizable AMOEBA model to demonstrate that these transferability errors can be reduced systemically and substantially without cross-terms by incorporating ion-ligand interactions during the parameterization of the protein force field. We first recalibrate the polarization descriptors of N-methylacetamide (NMA), which reduces errors in its high field response. Multi-parameter optimization using a Nelder-Mead-based algorithm of non-bonded interactions reproduces NMA’s reference data in both gas and condensed phases. Describing protein backbone, Gln/Asn side chains and termini caps using these optimized NMA parameters improves interactions of proteins with monovalent cations. However, a better agreement is achieved for an alternative NMA parameter set, which does not perform as well against condensed phase properties of NMA. Further improvement is achieved when Tyr hydroxyls are described using recalibrated methanol parameters whereas Ser and Thr side chains are described using recalibrated ethanol parameters. Overall, this recalibration reduces systematic errors in Na+-protein interaction energies from 6.0 to 2.5 kcal/mol and K+-protein interaction energy errors from 6.6 to 4.1 kcal/mol. These modifications do not affect AMOEBA’s reliability in predicting protein structure and dynamics. Analysis of the few cases for which ion-protein interaction energy errors remain substantially underestimated suggests paths forward for further improvements.