Adaptive Sampling using a Geometric Brownian Motion Model to Predict MD Trajectory Mobility on a Free Energy Surface
Adaptive Sampling using a Geometric Brownian Motion Model to Predict MD Trajectory Mobility on a Free Energy Surface
复制标题
使用几何布朗运动模型的自适应采样来预测自由能表面上的 MD 轨迹迁移率
DOI：
10.1016/j.bpj.2020.11.690
复制
发表时间：
2021
影响因子：
3.4
通讯作者：
Weinstein, Harel
中科院分区：
文献类型：
--
作者：
Kots, Ekaterina D.;Shore, Derek M.;Weinstein, Harel
提出了一种基于几何布朗运动数学模型的分子动力学（MD）轨迹群增强自适应采样方法。自适应采样协议通常从多个MD轨迹中提取感兴趣的集体变量（CV）的时间演化来确定MD群模拟中轨迹的方向。使用降维投影到诸如由时间无关分量分析（tICA）定义的空间上，然后将数据投影到由两个向量组成的2D空间上，其中两个向量沿CV沿着运动最慢。然后，从在该2D空间中识别的与初始构象最远的状态重新启动轨迹。这种传播方法并不能防止系统陷入局部最小值，或者回到初始状态，而不是在重新启动的轨迹中探索新的构象空间。考虑到在大型生化系统中显著的构象转变可能是罕见的事件，这些是重要的缺点。我们已经解决了这个问题，使该方法来预测未来的流动性的轨迹从他们以前的运动在tICA空间。由于自由能表面上的运动通常被认为是一个扩散过程，我们通过应用几何布朗运动数学模型（GBM）耦合到隐马尔可夫模型（HMM）为基础的事件检测算法来实现这一预测。我们提出了这种方法对不同大小的各种蛋白质系统的测试结果，包括膜蛋白（例如，多巴胺转运蛋白hDAT）。我们发现，这种方法提高了配置采样的效率，在MD模拟与预测精度为移动的和非移动的轨迹的85%和81%，分别。
We present an enhanced adaptive sampling method for swarms of Molecular Dynamics (MD) trajectories based on a mathematical model of Geometric Brownian Motion. Adaptive sampling protocols often determine the direction of trajectories in MD swarm simulation from the time evolution of collective variables (CVs) of interest extracted from multiple MD trajectories. Using dimensionality reduction projections such onto spaces such as defined by time-independent component analysis (tICA), the data are then projected onto a 2D space composed of two vectors with the slowest motion along the CVs. The trajectories are then re-initiated from the states identified in that 2D space that are the most distant from the initial conformation. This propagation approach does not prevent the system from being trapped in the local minimum, or from falling back into the initial state rather than exploring new conformational space in the re-initiated trajectories. These are important drawbacks considering that significant conformational transitions are likely rare events in large biochemical systems. We have addressed this issue by enabling the method to predict the future mobility of the trajectories from their previous motion on the tICA space. As movement on the free energy surface is commonly considered a diffusion process, we achieve this prediction by applying the geometric Brownian motion mathematical model (GBM) coupled to a Hidden Markov Model (HMM)-based event detection algorithm. We present the results from tests of this method on a variety of protein systems of different sizes, including membrane proteins (eg, the dopamine transporter hDAT). We found that this approach improves the efficiency of configurational sampling in MD simulations with a prediction accuracy for mobile and immobile trajectories of∼ 85% and∼ 81%, respectively.