Abstract The dynamic process of protein folding and conformation switching describes the basis of protein functions. Molecular dynamics (MD) simulations are precise computational tools for exploring protein dynamics, but the high computational costs make it difficult to scale up. Deep learning methods have been used to model the Boltzmann distribution of molecular simulations, but achieving MD-level accuracy remains a major challenge. Here, we present DynaFold, a generative deep learning framework based on latent diffusion for sampling protein dynamic trajectories. DynaFold accepts an initial structure and generalizes the conformational dynamics of different proteins with minimal trajectory data during training. It achieves state-of-the-art accuracy in predicting conformational ensembles and sampling conformational transition pathways, demonstrating superior generalization capability and computational efficiency compared to existing methods. Our framework provides a general solution for generating conformation distributions and transition processes between different conformations for proteins, enabling rapid sampling of structural ensembles and analysis of Boltzmann systems.
Chen et al. (Thu,) studied this question.