AI 中文总结
该研究提出自洽超粗粒化（SC-UCG）方法，通过图消息传递等技术解决超粗粒化忽略内部态关联的问题，在四聚体二阶相变模拟中，仅用单温度数据集训练即可准确捕获内部态切换与相变。
AI 中文摘要
自底向上的粗粒化方法可拓展分子动力学（MD）模拟可及的长度与时间尺度，但信息丢失会阻碍复杂生物分子动力学中多态现象的准确表征。超粗粒化（UCG）将离散的“类量子”扩展自由度（即“内部态”）投影到粗粒化（CG）分子上，提升了CG模型的表达能力。UCG中的快速局部平衡（RLE）近似依赖用户定义的集体变量（CV，如局部密度），且忽略CG分子内部及分子间的内部态关联。本文提出自洽UCG（SC-UCG），该方法直接利用底层UCG相互作用分配内部态，无需在CG系综中设计CV；模拟过程中，内部态概率通过图消息传递自洽确定。我们用Bethe近似与基于AI的推断增强RLE哈密顿量，以表征UCG珠子间的显式关联。针对力场训练，我们开发了多层内部态一致性（MISC）——一种基于相对熵最小化的机器学习方法，可避免中间力场的迭代采样。我们将SC-UCG应用于一种四聚体，该四聚体呈现从超临界外消旋流体到亚临界富D与富L流体的二阶对称性破缺相变；尽管仅用单温度数据集训练，SC-UCG仍能捕获亚临界区域内部态的集体切换，并重现跨温度的相变。
英文摘要
Bottom-up coarse-graining expands the length and time scales accessible to molecular dynamics (MD) simulations, but information loss can hinder accurate representation of multistate phenomena in complex biomolecular dynamics. Ultra-Coarse-Graining (UCG) projects discrete "quantum-like" extended degrees of freedom, or "internal states," onto coarse-grained (CG) molecules, extending CG model expressiveness. The rapid-local-equilibrium (RLE) approximation in UCG depends on user-defined collective variables (CVs, e.g., local density) and neglects correlations between internal states within and between CG molecules. We present Self-Consistent UCG (SC-UCG), which uses the underlying UCG interactions directly to assign internal states without designing CVs in the CG ensemble. During simulation, internal state probabilities are determined self-consistently through graph message passing. We enhance the RLE Hamiltonian with the Bethe approximation and AI-based inference to represent explicit correlations between UCG beads. For force-field training, we develop Multilayer Internal State Consistency (MISC), a machine-learning method derived from relative entropy minimization that avoids iterative sampling of intermediate force fields. We apply SC-UCG to a tetramer exhibiting a second-order symmetry-breaking phase transition from a supercritical racemic fluid to subcritical D-rich and L-rich fluids. SC-UCG captures collective switching of internal states in the subcritical region and recapitulates the phase transition across temperatures, despite being trained on a single-temperature dataset.
