Efficient Molecular Dynamics simulations using 3D Graph Neural Networks

Molecular dynamics (MD) simulations are an important tool in many areas of science as they can simulate chemical processes using Newtonian dynamics or quantum mechanics. However, as they require the steps to be calculated every ~1fs, they are slow and currently unsuitable for simulating processes at long, biologically-relevant time scales. Their prohibitively large computational cost prevents them from being used on a large scale for tasks such as protein folding, molecular docking or assessment of ligand stability in a protein pocket. The aim of the project is to develop computational methods to obtain large time-step molecular dynamics simulations using machine learning techniques. The idea of achieving MD speed-ups has been present since the creation of this computational tool. However, the current approaches often suffer from poor system stability over long simulation runs, do not account for physical symmetries making them less data-efficient, or are not transferable between chemical systems so cannot be reused on a different molecule after training.
One of the project's goals is to utilise 3D Graph Neural Networks (GNNs) which can represent the molecular systems as graphs embedded in 3D Euclidian space and account for the system's rotational and translational symmetries. Such chosen methodology should, in principle, be able to overcome the challenges of transferability or data efficiency. Such architecture will be then combined with generative AI approaches, such as diffusion or flows, to obtain large-time-step molecular dynamics trajectories. The approach will be tested on various objectives such as the stability of predicted trajectories over time, the ability to explore different regions of the potential energy landscape and, ideally, the ability to minimise the systems energy. Such system could be then used to explore chemical processes happening near the energy equilibrium such as generation of molecular conformations or processes based on energy minimization such as protein folding.
The approach chosen for this project will require careful consideration from the theoretical site to best design the graph neural networks for this task. On top of that, the research will involve coming up with new probabilistic and generative frameworks to be able to model the stochastic nature of the chemical dynamics with the desired accuracy and properties. Therefore, this project falls within the EPSRC Artificial Intelligence Technologies, Computational and Theoretical Chemistry, and Statistics and Applied Probability research areas.