Training Set for the Prediction of Partial Charges

Abstract

Parametrization of small organic molecules for classical molecular dynamics simulations is not trivial. The vastness of the chemical space makes approaches using building blocks challenging. The most common approach is therefore an invidual parametriza- tion of each compound by deriving partial charges from semi-empirical or ab initio calculations and inheriting the bonded and van der Waals (Lennard-Jones) parameters from a biomolecular force field. The quality of the partial charges generated in this fashion depends on the level of the quantum-chemical calculation as well as on the extraction procedure used. Here, we present a machine learning (ML) based approach for predicting partial charges extracted from density functional theory (DFT) electron densities. The training set was chosen with the goal to provide a broad coverage of the known chemical space of drug-like molecules. In addition to the speed of the approach, the partial charges predicted by ML are not dependent on the molecular conformation in contrast to the ones obtained by fitting to the electrostatic potential (ESP).