Assessments of Variational Autoencoder in Protein Conformation Exploration

Molecular dynamics (MD) simulations have been extensively used to study protein dynamics and subsequently functions. However, they are unable to sufficiently explore the conformational space within equilibrium timescales. The purpose of this study is to evaluate the feasibility of using variational autoencoders to assist the exploration of protein conformational landscapes. Using three modeling systems, we show that variational autoencoders are capable of capturing high-level hidden information which distinguishes alternate protein conformations, which can be readily used for generating unseen and physically plausible protein conformations to direct sampling to favored conformational spaces. Based on our investigations, we also find that VAE prefers interpolation than extrapolation and increasing latent space dimension can lead to a trade-off between performances and difficulties, thus we propose that the initial data preparation is important for building VAE models to explore protein conformational landscapes.