Langevin Modeling of Rare Biomolecular Processes

对罕见生物分子事件的电子预测是一个长期存在的问题，因为缓慢的(比方说，毫秒级的)过程通常超出了当前全原子分子动力学(MD)模拟的范围。为了避免这种时间尺度问题，人们经常调用大规模并行计算策略或各种增强的采样技术，例如副本交换、引导或目标MD。虽然这些方法可能会提供系统自由能图景的有效采样，但动力学描述需要一些粗粒度的“后仿真”模型，该模型能够从采样数据重建动力学。为此，我们最近提出了一种数据驱动的朗之万方程(Del)方法，该方法根据给定的MD轨迹构建低维动力学模型。现在，挑战是利用LED公式的优点来促进其在从多肽到多结构域蛋白质的各种生物分子系统中的实际应用。在这里，我们将遵循两条可行而有希望的路线：(1)利用温度作为驱动力来克服能量障碍，我们将通过副本交换MD模拟来构建一个DELE模型。此外，我们将利用涨落耗散定理从高温模拟中进行低温预测。(2)采用元动力学和定向或定向MD等增强采样方法，对整个自由能格局进行有效采样，并为随后的短MD轨迹选择多个起始构型，这些初始构型为DELD模型生成输入数据。采用不断增加的大小和动力学复杂性的蛋白质系统，我们最终将考虑我们的实验合作者所研究的系统，并旨在比较计算和实验对分子结构和构象转变速率的预测。

The in silico prediction of rare biomolecular events represents a long-standing problem, because slow (say, millisecond timescale) processes are typically out of reach of current all-atom molecular dynamics (MD) simulations. To circumvent this time scale problem, one often invokes massive parallel computing strategies or various enhanced sampling techniques such as replica exchange, steered or targeted MD. While these approaches may provide an efficient sampling of the system's free energy landscape, a dynamical description requires some coarse-grained "post-simulation'' model that is capable of rebuilding the kinetics from the sampled data. To this end, we have recently proposed a data-driven Langevin equation (dLE) approach that constructs a low-dimensional dynamical model from a given MD trajectory. Now, the challenge is to exploit the virtues of the dLE formulation to facilitate its practical application to various biomolecular systems, ranging from polypeptides to multidomain proteins. Here two viable and promising routes will be followed: (1) Using temperature as driving force to overcome energy barriers, we will construct a dLE model from replica exchange MD simulations. Moreover, we will exploit a fluctuation-dissipation theorem in order to make low-temperature predictions from high-temperature simulations. (2) Employing enhanced sampling methods such as metadynamics and steered or targeted MD, we perform an efficient sampling of the overall free energy landscape and choose a number of starting configurations for subsequent short MD trajectories, which generate input data for the dLE model. Adopting protein systems of increasing size and dynamical complexity, we will ultimately consider the systems studied by our experimental collaborators and aim to compare computational and experimental predictions of molecular structures and conformational transition rates.