Abstract
Simulating large proteins using traditional molecular dynamics (MD) is computationally demanding. To address this challenge, we propose a novel tree-structured coarse-grained model that efficiently captures protein dynamics. By leveraging a hierarchical protein representation, our model accurately reconstructs high-resolution protein structures, with sub-angstrom precision achieved for a 168-amino acid protein. We combine this coarse-grained model with a deep learning framework based on stochastic differential equations (SDEs). A neural network is trained to model the drift force, while a RealNVP-based noise generator approximates the stochastic component. This approach enables a significant speedup of over 20,000 times compared to traditional MD, allowing for the generation of microsecond-long trajectories within a few minutes and providing valuable insights into protein behavior. Our method demonstrates high accuracy, achieving sub-angstrom reconstruction for short (25 ns) trajectories and maintaining statistical consistency across multiple independent simulations.