PhD defence by Marloes Elisabeth Arts
Summary
Firstly, we propose a general method to simultaneously impose local and global constraints in protein structure ensemble modelling. As a proof of principle, this method is incorporated into a simple variational autoencoder (VAE) and we demonstrate that the generated samples are of high quality, both locally and globally. The second contribution is a denoising diffusion model based method trained on reduced representations of protein structures from molecular dynamics simulations. Not only can this model produce new samples in a one-shot manner, a force field can also cheaply be extracted to perform new simulations. Specifically, we examine the strengths and weaknesses of Bayesian decoders in this context as well as show the potential of hierarchical VAEs to alleviate the mismatch between the commonly used standard Gaussian prior over latent space and the ``star-shaped aggregated posterior for protein family data.