利用回归模型筛选出近天然的抗原-抗体对接模拟结构

Regression analysis to select native-like structures from decoys of antigen-antibody docking

摘要:

在抗原-抗体分子对接模拟所生成的大量计算生成构象中筛选出近天然结构，即接近真实情况的抗原-抗体结合模式。借鉴QSAR原理，定义抗原-抗体接触面描述符并利用Discovery Studio 4.5软件平台计算出各对接模拟构象的接触面描述符和能量参数。构造训练集数据进行回归分析，建立预测对接模拟构象是否是近天然结构的数学模型。通过测试集和实际应用情况检验该数学模型。通过回归分析所建立的数学模型能够在成百上千的抗原-抗体对接模拟构象中有效筛选出其中的近天然结构，在测试集验证和4G7抗体结合模式预测应用中具有良好的表现，验证了该数学模型的有效性和实用性。经验性的抗原-抗体接触面特征如氢键密度、氨基酸对偏好性指数等以及能量参数能够共同有效表征近天然结构，所建立的数学模型有效增强了通过分子对接预测抗原-抗体结合模式的可行性。

Abstract:

Given the increasing exploitation of antibodies in different contexts such as molecular diagnostics and therapeutics, it would be beneficial to unravel properties of antigen-antibody interaction with modeling of computational protein-protein docking, especially, in the absence of a cocrystal structure. However, obtaining a native-like antigen-antibody structure remains challenging due in part to failing to reliably discriminate accurate from inaccurate structures among tens of thousands of decoys after computational docking with existing scoring function. We hypothesized that some important physicochemical and energetic features could be used to describe antigen-antibody interfaces and identify native-like antigen-antibody structure. We prepared a dataset, a subset of Protein-Protein Docking Benchmark Version 4.0, comprising 37 nonredundant 3D structures of antigen-antibody complexes, and used it to train and test multivariate logistic regression equation which took several important physicochemical and energetic features of decoys as dependent variables. Our results indicate that the ability to identify native-like structures of our method is superior to ZRANK and ZDOCK score for the subset of antigen-antibody complexes. And then, we use our method in workflow of predicting epitope of anti-Ebola glycoprotein monoclonal antibody—4G7 and identify three accurate residues in its epitope.