Robust and Automated Force Field Parameterization Using Validation Sets and Active Learning

Molecular mechanics force fields enable atomistic simulations of complex systems that are too large for a quantum mechanical treatment. Simulation accuracy depends on the parameters employed in the force field. Every new molecule must have parameters generated for it, either by using a general force field or fitting a custom parameter set for that system. While fitting custom parameter sets can provide superior accuracy compared to a general force field, the process of single-molecule force field fitting is often tedious, expensive, and bespoke. We present an automated and iterative procedure for fitting single-molecule force fields. This program optimizes the parameters with respect to a data set of quantum mechanical (QM) calculations, runs dynamics with the new parameters to sample new conformations, computes QM energies and forces on those conformations, adds them to the data set, and returns to the parameter optimization step. In contrast to previous attempts at iterative optimization, we employ a validation set to determine convergence. Using a validation set circumvents problems with parameter convergence and flags when overfitting occurs. As an example, we find that Boltzmann sampling at 400 K is sufficient to fit a force field for a trialanine peptide, a system with a rugged potential energy surface. Last, we demonstrate the efficiency of the method by fitting a custom force field for each molecule in a library of 31 photosynthesis cofactors.分子力学力场（Molecular mechanics force fields）可实现对体积过于庞大、无法通过量子力学处理的复杂体系开展原子级模拟。模拟精度取决于力场中所采用的参数。每一种新分子都需要为其生成适配参数，既可以通过通用力场实现，也可以针对该体系拟合定制化参数集。尽管相较于通用力场，拟合定制化参数集能够提供更优异的模拟精度，但单分子力场的拟合过程往往繁琐耗时、成本高昂且高度定制化。我们提出了一种自动化且可迭代的单分子力场拟合流程。该程序针对量子力学（Quantum Mechanical, QM）计算数据集对参数进行优化，使用新参数运行分子动力学模拟以采样新的构象，计算这些构象上的QM能量与作用力，将其添加至数据集后，返回至参数优化步骤。与此前的迭代优化尝试不同，我们采用验证集来判定收敛性。使用验证集可规避参数收敛相关问题，并在出现过拟合时发出提示。作为示例，我们发现400 K下的玻尔兹曼采样足以拟合三丙氨酸肽的力场，该体系具有崎岖的势能面。最后，我们通过为包含31种光合作用辅因子的分子库中的每一种分子拟合定制化力场，验证了该方法的高效性。