Creating knowledgeable atoms for the molecular dynamics simulations of peptides.

Important diseases such as Alzheimer's have a molecular cause, which needs to be understood at fundamental level in order to make solid progress in designing a potential cure. It was discovered that Intrinsically Disordered Proteins/peptides (IDPs) underpin Alzheimer's disease. This disorder challenges the standard tools of structural biology, so an independent source of information [1], i.e. computational molecular dynamics is needed.
Behind any molecular dynamics simulation is an energy prediction function, which cannot be obtained from first-principles due to computational expense. Hence, force fields are used, which must return a reliable system energy. Unfortunately, traditional force field architecture does not achieve this, which is why it was overhauled by our new force field called FFLUX [2]. FFLUX abandons old approximations and enhances the realism of the energy prediction. Key to FFLUX is the use of a machine learning method called Gaussian Processes (GP) inference [3]. This method successfully predicts atomic properties (energies, multipole moments) [4] directly from the positions of surrounding atoms, based on a sufficient number of training geometries.
FFLUX has a better prospect to make the right predictions, for the right reasons, thanks to the use of machine learning because it perceives the structure of peptides as an interplay of four fundamental energy contributions at atomic level: (i) the intra-atomic self-energy, responsible for stereo-electronic effects, the inter-atomic (ii) electrostatic energy, capturing charge transfer and polar effects, (iii) the exchange energy, modelling delocalisation, hydrogen bonding and bond strength, while (iv) dynamic correlation takes care of dispersion. As such, FFLUX also provides deeper insight into peptide conformation and aggregation, beyond that of a traditional force field. As such it is better placed to make solid progress in the atomistic understanding of the nucleation process behind Alzheimer's disease.
In this project, the GP inference methods within FFLUX will be improved to incorporate recent advances in computational inference methodology. Bayesian optimisation techniques will be used for actively selecting the best function evaluations to approximate the energy function in a range of scenarios, e.g. for both off-line (developing the energy function prior to application) and on-line (further improving the energy function within a specific application) applications. Recent advances in GP inference, e.g. as implemented in the recently developed GPFlow package [5] that leverages TensorFlow, will allow improved scalability, particularly for non-Gaussian data likelihoods. Through these advances the student will help create the next generation FFLUX package.