An information theory-based machine learning approach to detecting functionally conserved and coordinated protein dynamics

The application of machine learning classification to the molecular dynamics of the functional states of protein allows for application of an information theoretic framework familiar to traditional bioinformatics. The functional states of proteins involving binding interactions with partners comprised of protein, DNA or small molecules can first be defined in a binary fashion (i.e. bound vs unbound), subsequently simulated in molecular dynamics software, and then employed as a comparative training set for a binary machine learning classifier capable of discerning the complex dynamical consequences of binding interaction. This learner can subsequently be deployed on new simulations of the functionally bound state to validate its ability to recognize the molecular motions that are supporting binding function. Regions of proteins with functionally conserved dynamics will induce significant local correlations in learning performance across independent validation runs. Through case studies of Rbp subunit 4/7 interaction in RNA Pol II and DNA-protein interactions of TATA binding protein, we demonstrate this method of detecting functionally conserved protein dynamics. We also demonstrate how Shannon information, relative entropy and mutual information can be applied to these binary classification states of dynamic simulations in order to compare dynamics and identify concerted motions involved in dynamic interactions across sites.将机器学习分类方法应用于蛋白质功能状态的分子动力学研究，可引入传统生物信息学中常用的信息论分析框架。涉及与由蛋白质、DNA或小分子构成的结合伴侣发生相互作用的蛋白质功能状态，可首先以二元方式定义（即结合态与未结合态），随后通过分子动力学软件完成模拟，之后可作为二元机器学习分类器的对比训练集——该分类器能够识别结合相互作用所引发的复杂动力学后果。随后可将训练好的分类器应用于功能结合态的新模拟数据中，以验证其识别支撑结合功能的分子运动的能力。具有功能保守动力学特征的蛋白质区域，会在多次独立验证运行的学习性能中产生显著的局部相关性。通过针对RNA聚合酶II（RNA Pol II）中Rbp亚基4/7相互作用与TATA结合蛋白的DNA-蛋白质相互作用开展的案例研究，我们证实了该检测功能保守蛋白质动力学的方法的有效性。我们还展示了如何将香农信息、相对熵与互信息应用于动态模拟的二元分类状态，以对比动力学特征并识别参与跨位点动态相互作用的协同运动。