For every retinylidene a force field: Using machine learning to improve molecular simulations of rhodopsins.

Abstract

G-protein-coupled receptors (GPCRs) are a superfamily of transmembrane proteins that perform diverse signaling roles in organisms, including humans. This makes them an attractive target for pharmaceuticals, with 134 different GPCRs being currently targeted by roughly 700 drugs. Here we study the retinylidene cofactor of GPCR rhodopsin that is responsible for initiating phototransduction. There are many distinct retinal forms—isomers, rotamers, and protonation states—that all have unique electronic landscapes. However, current molecular dynamics force field libraries are incomplete in that they ignore these differences and treat each retinylidene form identically. It is unknown what effect this has on current molecular dynamics simulations of retinal proteins. Our work addresses this problem by using machine learning methods to generate unique force fields for each of forty retinylidene forms. This approach improved on traditional parameterization schemes by using Born-Oppenheimer molecular dynamics trajectories to sample the retinylidene energetic landscapes. To validate our methods, the new force fields were then used in all-atom molecular dynamics simulations of the entire rhodopsin-membrane-water system. From these simulations we were able to recover hydrogen relaxation rates that closely matched experimental solid-state deuterium (2H) NMR measurements. Comparison of force field parameters showed a small but—given the highly allosteric nature of rhodopsin—potentially significant difference between retinylidene forms. We expect this work will improve simulations of many different types of rhodopsins, and that it will lead to an increased appreciation of the complexity of the retinylidene energetic landscape. Work supported by US NIH (R01EY026041) and US NSF (CHE 1904125, MCB 1817862) and Ministry of Education and Science of the Russian Federation (Project FSRM-2020-0006).