Toward Prediction of Electrostatic Parameters for Force Fields That Explicitly Treat Electronic Polarization
Toward Prediction of Electrostatic Parameters for Force Fields That Explicitly Treat Electronic Polarization
复制标题
DOI：
10.1021/acs.jctc.8b01289
复制
发表时间：
2019-04-01
影响因子：
5.5
通讯作者：
MacKerell, Alexander D., Jr.
中科院分区：
文献类型：
--
作者：
Heid, Esther;Fleck, Markus;MacKerell, Alexander D., Jr.
发展极化力场的原子极化率的推导一直是一个长期存在的问题。原子的极化率通常是从表格值开始手动改进的，这使得参数的自动分配变得困难，并阻碍了所获得值的重复性和可转移性。为了克服这一点，我们对大量高质量的量子力学原子极化率和部分原子电荷训练了线性增量方案和多层感知器神经网络，其中只使用每个原子的类型及其连接性作为输入。用神经网络预测原子的极化率和电荷的平均误差为0.023埃(3)和0.019埃，用简单增量法预测的原子极化率和电荷的平均误差为0.063埃(3)和0.069埃。由于该算法只依赖于分子内原子的连接性，从而省略了对三维构象的依赖，该方法自然地将相似的电荷和极化率分配给对称基团。因此，根据极化力场的发展需要，提出了一种产生有机分子的部分原子电荷和原子极化率的方便实用方法。
The derivation of atomic polarizabilities for polarizable force field development has been a long-standing problem. Atomic polarizabilities were often refined manually starting from tabulated values, rendering an automated assignment of parameters difficult and hampering reproducibility and transferability of the obtained values. To overcome this, we trained both a linear increment scheme and a multilayer perceptron neural network on a large number of high-quality quantum mechanical atomic polarizabilities and partial atomic charges, where only the type of each atom and its connectivity were used as input. The predicted atomic polarizabilities and charges had average errors of 0.023 angstrom(3) and 0.019 e using the neural net and 0,063 angstrom(3) and 0.069 e using the simple increment scheme. As the algorithm relies only on the connectivities of the atoms within a molecule, thus omitting dependencies on the three-dimensional conformation, the approach naturally assigns like charges and polarizabilities to symmetrical groups. Accordingly, a convenient utility is presented for generating the partial atomic charges and atomic polarizabilities for organic molecules as needed in polarizable force field development.