Using Neural Networks to Accelerate Molecular Dynamics

When running molecular dynamics simulations, typically timesteps must be on the order of 2fs to preserve numerical stability. This sharply limits our ability to generate trajectories for processes such as protein folding that can take on the order of milliseconds. In this talk, we'll discuss training a neural net to predict the configuration of the protein many timesteps in the future (conditional on its current configuration) in order to save computation. Such a neural net cannot output a state deterministically but must sample from a probability distribution because molecular dynamics is stochastic. We'll examine two well-known machine learning methods for modelling a probability distribution over many dimensions: Generative Adversarial Nets and Denoising Diffusion Probabilistic Models. In addition, molecular dynamics is symmetric under rotations and translations, and so the talk will also explain how the idea of a representation from group theory is helpful for constructing neural nets that respect these symmetries. We try out these techniques on small molecules and short proteins in solution.