Abstract

The aim of my lab is to develop accurate, transferable and fast force fields for biomolecular simulation. I will describe two projects. Firstly, the development of an implicit solvent model using the technique of differentiable simulation to better describe intrinsically disordered proteins (IDPs). This force field, GB99dms, fixes the over-compaction of IDPs in implicit solvent and allows fast conformational exploration. Secondly, the training of an all-atom biomolecular force field without reference to previous parameters. This model, Garnet, uses a graph neural network for continuous atom typing and is trained on quantum mechanical, condensed phase and protein NMR data. The force field shows comparable performance to existing, manually-parameterised force fields and is competitive at predicting binding free energies across a range of targets.