Transferable kriging machine learning models for the multipolar electrostatics of helical deca-alanine
Transferable kriging machine learning models for the multipolar electrostatics of helical deca-alanine
复制标题
DOI：
10.1007/s00214-015-1739-y
复制
发表时间：
2015-10-17
影响因子：
1.7
通讯作者：
Popelier, Paul L. A.
中科院分区：
文献类型：
--
作者：
Fletcher, Timothy L.;Popelier, Paul L. A.
我们利用量子拓扑原子的可迁移性来构造多极可极化蛋白质力场QCTFF。用克立格机器学习方法预测了一个由10个丙氨酸残基(103个原子)组成的螺旋寡肽的总静电能，平均误差为6.4kJ·mol(-1)。这个错误类似于在过去的QCTFF出版物中发现的小分子中的错误。克立格法将分子几何与描述从头算电子密度的原子多极矩联系起来。原子类型由螺旋内类似的原子构成。由于给定原子类型中的原子共享局部化学环境，它们可以共享具有更少数量的输入描述符(即特征)的克里金模型。特征约简使克立格训练次数减少了23倍以上，但预测误差仅增加了1.3%。在可转移性测试中，可转移模型在预测训练集外的原子的矩时给出了5.7%的误差，而当针对属于训练数据中包括的原子的数据进行测试时误差为3.9%。可转移的克立格模型以有用的精度成功地预测了原子多极矩，为整个蛋白质的QCTFF建模开辟了一条道路。
We exploit the transferability of quantum topological atoms in the construction of a multipolar polarizable protein force field QCTFF. A helical oligopeptide of 10 alanine residues (103 atoms) has its total electrostatic energy predicted using the kriging machine learning method with a mean error of 6.4 kJ mol(-1). This error is similar to that found in smaller molecules presented in past QCTFF publications. Kriging relates the molecular geometry to atomic multipole moments that describe the ab initio electron density. Atom types are constructed from similar atoms within the helix. As the atoms within a given atom type share a local chemical environment, they can share a kriging model with a reduced number of input descriptors (i.e. features). The feature reduction decreases the kriging training times by more than 23 times but increases the prediction error by only 1.3 %. In transferability tests, transferable models give a 5.7 % error when predicting moments of an atom outside the training set, compared to the 3.9 % error when tested against data belonging to atoms included in the training data. The transferable kriging models successfully predict atomic multipole moments with useful accuracy, opening an avenue to QCTFF modelling of a whole protein.