Developing bespoke force fields and associated toolkits for computer aided drug design

Abstract

The ability to create accurate models of molecules enables the calculation of key properties, and the simulation of interactions. One such model is the force field. In this thesis I further develop the QUBE force field, based on the OPLS force field, and its associated toolkit, QUBEKit. QUBEKit is a piece of Python-based software which makes use of quantum mechanical calculations to derive molecular mechanics parameters for simulation rather than building transferable libraries of parameters. Its development has allowed fast, automated generation of force field parameters for property calculations. Within force fields, charges are typically treated as atom-centred points, approximating the electrostatics of the molecule as a whole. In this thesis, I not only show that this approximation can be highly inaccurate, but I also introduce a method of improvement through the use of virtual sites. The fitting of these extra charges around certain atoms leads to improved modelling of charge anisotropy, which in turn improved the accuracy of liquid properties in simulation. Often in force field design, hyperparameters such as the implicit solvent dielectric or the basis set (both used for quantum mechanical calculations) are baked into the production of the force field parameters. These hyperparameters can be somewhat arbitrarily chosen and hard to analyse and test, yet can impact the final accuracy heavily. We introduce a means of automatically comparing and optimising these hyperparameters to ultimately improve the accuracy of various molecular properties.