Prediction of zwitterion hydration and ion association properties using machine learning

Abstract

Molecular dynamics simulations were performed to study the hydration and ion association properties of a library of zwitterionic molecules with varying charged moieties and spacer chemistries in pure water and with Na⁺ and Cl⁻ ions. The structure and dynamics of associations were calculated using the radial distribution and residence time correlation function. Resulting association properties are used as target variables for a machine learning model, with cheminformatic descriptors of the molecule subunits used as descriptors. Prediction of hydration properties revealed that steric and hydrogen bonding descriptors were of greatest importance and there was influence from the cationic moiety on the anionic moiety hydration properties. Ion association properties prediction performed poorly, which is attributed to the role of hydration layers in ion association dynamics. This study is the first to quantitatively describe the influence of subunit chemistry on hydration and ion association properties of zwitterions. These quantitative descriptions supplement prior studies of zwitterion association and previously described design principles.