Predicting the conformations adopted by flexible molecules is an important feature in drug design because efficacy strongly depends on the complementary arrangements of functional groups in the drug and the target. Molecular dynamics simulations can be used to predict the conformational preferences of flexible drug molecules, but the failings of molecular mechanics (MM) potentials for this task are well documented. In this study, we present a protocol for the development of machine learned potentials (MLPs) for accurate prediction of drug molecule conformations in solution using molecular dynamics simulations.

We focus our attention on molecules containing the phenylethylamine scaffold, which span a large chemical space and have numerous therapeutic applications, including the treatment of neurological disorders, cardiovascular disease and urinary systems. The conformational landscape of these compounds is dominated by rotation around the C_α-C_β bond, which in turn depends on a subtle balance between intramolecular and solvent interactions, as well as the protonation state of the amine.

Molecule-specific MLPs were trained according to the recently developed PairFENet scheme, based on the original PairFNet scheme. The PairFENet scheme trains MLPs that are energy-conserving, invariant to rotations, translations and (effectively, despite the use of global molecular descriptors) atomic permutations and able to accurately predict force and energy (MAEs << 1 kcal/mol(/Å). With respect to prediction of torsional energy profiles for rotation about the C_α-C_β bond crucial for accurate assessment of conformational stability, we show that MLPs trained according to PairFENet outperform both MM potentials and other MLP schemes that use local molecular descriptors (e.g. ANI).

Solution phase simulations of phenylethylamines using PairFENet-trained MLPs necessitate the use of a hybrid scheme in which solvent molecules are modelled using MM potentials. To enable such simulations, PairFENet was modified to predict conformation-dependent partial atomic charges in addition to atomic forces and energies. Finally, we calculate conformational free energy surfaces using enhanced sampling and discuss the impact of the fluctuating charges on predicted conformations.

Neil Burton

Appearances