Biomolecules, such as enzymes, often undergo large-scale conformational changes that are essential to their biological functions. While conventional molecular dynamics (MD) simulations provide critical insights into the dynamics at equilibrium states, elucidating the detailed transition pathways between these states remains a significant challenge. Established methods like linear morphing can visualize a plausible transition with low computational cost, but their underlying linear interpolation makes it difficult to capture the specific sequence of domain motions and complex side-chain rearrangements. To address these limitations, we have developed MOVE-DM v.3.0, an AI-based, all-atom nonlinear morphing method. This framework uniquely integrates molecular simulations with deep learning to generate physically realistic transition pathways. The method begins by leveraging short MD simulations of the initial and final conformational states. An initial pathway is then estimated on a potential of mean force (PMF) surface using a string-of-beads model. Subsequently, a deep neural network (DNN) is trained on the structural data from these endpoint simulations. The trained DNN then generates a refined, all-atom morphing trajectory that incorporates the intricate motions of side-chains throughout the transition. MOVE-DM v.3.0 can capture both the sequential order of domain-level structural changes and detailed side-chain rearrangements, providing a comprehensive view of the conformational transition. By combining the strengths of MD simulation and AI, our method enables the reconstruction of pathways at a level of detail unachievable with conventional approaches. This practical framework opens new avenues for analyzing complex biomolecular dynamics and understanding the functional mechanisms of proteins at atomic resolution.
Miyashita et al. (Sun,) studied this question.