Fenn T D, Schnieders M J
Department of Bioengineering, Stanford University, Stanford, California, USA.
Acta Crystallogr D Biol Crystallogr. 2011 Nov;67(Pt 11):957-65. doi: 10.1107/S0907444911039060. Epub 2011 Oct 19.
In the past, weighting between the sum of chemical and data-based targets in macromolecular crystallographic refinement was based on comparing the gradients or Hessian diagonal terms of the two potential functions. Here, limitations of this scheme are demonstrated, especially in the context of a maximum-likelihood target that is inherently weighted by the model and data errors. In fact, the congruence between the maximum-likelihood target and a chemical potential based on polarizable atomic multipole electrostatics evaluated with Ewald summation has opened the door to a transferable static weight. An optimal static weight is derived from first principles and is demonstrated to be transferable across a broad range of data resolutions in the context of a recent implementation of X-ray crystallographic refinement using the polarizable AMOEBA force field and it is shown that the resulting models are balanced with respect to optimizing both R(free) and MolProbity scores. Conversely, the classical automatic weighting scheme is shown to lead to underfitting or overfitting of the data and poor model geometry. The benefits of this approach for low-resolution diffraction data, where the need for prior chemical information is of particular importance, are also highlighted. It is demonstrated that this method is transferable between low- and high-resolution maximum-likelihood-based crystallographic refinement, which proves for the first time that resolution-dependent parameterization of either the weight or the chemical potential is unnecessary.
过去，在大分子晶体学精修中，基于化学目标和基于数据目标的总和进行加权是通过比较两个势函数的梯度或海森矩阵对角项来实现的。在此，展示了该方案的局限性，特别是在由模型和数据误差固有加权的最大似然目标的背景下。事实上，最大似然目标与基于用埃瓦尔德求和评估的可极化原子多极静电学的化学势之间的一致性为可转移的静态权重打开了大门。从第一原理推导得出一个最优静态权重，并在最近使用可极化AMOEBA力场进行X射线晶体学精修的背景下，证明其在广泛的数据分辨率范围内是可转移的，结果表明所得模型在优化R(free)和MolProbity分数方面是平衡的。相反，经典的自动加权方案被证明会导致数据拟合不足或过度拟合以及模型几何结构不佳。还强调了这种方法对于低分辨率衍射数据的好处，在这种情况下，对先验化学信息的需求尤为重要。结果表明，该方法可在基于低分辨率和高分辨率最大似然的晶体学精修之间转移，这首次证明了权重或化学势的分辨率依赖参数化是不必要的。
