Salawu Emmanuel Oluwatobi
Machine Learning Solutions Lab, Amazon Web Services (AWS), Herndon, VA, United States.
Front Mol Biosci. 2021 May 4;8:587151. doi: 10.3389/fmolb.2021.587151. eCollection 2021.
The molecular structures (i.e., conformation spaces, CS) of bio-macromolecules and the dynamics that molecules exhibit are crucial to the understanding of the basis of many diseases and in the continuous attempts to retarget known drugs/medications, improve the efficacy of existing drugs, or develop novel drugs. These make a better understanding and the exploration of the CS of molecules a research hotspot. While it is generally easy to computationally explore the CS of small molecules (such as peptides and ligands), the exploration of the CS of a larger biomolecule beyond the local energy well and beyond the initial equilibrium structure of the molecule is generally nontrivial and can often be computationally prohibitive for molecules of considerable size. Therefore, research efforts in this area focus on the development of ways that systematically favor the sampling of new conformations while penalizing the resampling of previously sampled conformations. In this work, we present (DESP), a technique for enhanced sampling that combines molecular dynamics (MD) simulations and deep neural networks (DNNs), in which biasing potentials for guiding the MD simulations are derived from the KL divergence between the DNN-learned latent space vectors of [a] the most recently sampled conformation and those of [b] the previously sampled conformations. Overall, DESP efficiently samples wide CS and outperforms conventional MD simulations as well as accelerated MD simulations. We acknowledge that this is an actively evolving research area, and we continue to further develop the techniques presented here and their derivatives tailored at achieving DNN-enhanced steered MD simulations and DNN-enhanced targeted MD simulations.
生物大分子的分子结构（即构象空间，CS）以及分子所表现出的动力学对于理解许多疾病的基础以及不断尝试重新靶向已知药物、提高现有药物疗效或开发新型药物至关重要。这些使得对分子构象空间的更好理解和探索成为一个研究热点。虽然通常很容易通过计算探索小分子（如肽和配体）的构象空间，但对于超出局部能量阱以及分子初始平衡结构的更大生物分子的构象空间进行探索通常并非易事，而且对于相当大尺寸的分子，计算上往往是 prohibitive（此处原词有误，推测可能是“prohibitive”，意为“令人望而却步的”）。因此，该领域的研究工作集中在开发一些方法，这些方法在系统地有利于新构象采样的同时，对先前采样构象的重新采样进行惩罚。在这项工作中，我们提出了一种增强采样技术（DESP），它将分子动力学（MD）模拟和深度神经网络（DNN）相结合，其中用于指导MD模拟的偏置势是从最近采样构象的DNN学习潜在空间向量与先前采样构象的DNN学习潜在空间向量之间的KL散度推导出来的。总体而言，DESP能够有效地对广泛的构象空间进行采样，并且优于传统的MD模拟以及加速MD模拟。我们认识到这是一个不断发展的研究领域，并且我们将继续进一步开发此处介绍的技术及其衍生物，以实现DNN增强的导向MD模拟和DNN增强的靶向MD模拟。
