【摘要】 细胞信号网络中的蛋白质相互作用常通过结合-折叠偶合方式实现,即来自一方蛋白的刚性肽识别域与来自另一方蛋白表面的一段柔性寡肽片段发生识别和结合,从而介导母体蛋白相互作用.深入分析域/肽识别和相互作用的理化性质及精确预测其作用行为,能够有效揭示细胞信号转导的分子基础.该研究将一种新型非线性机器学习方法即高斯过程(GP),用于预测和分析4类域/肽体系数千个样本的亲和力数值和序列结构特征,并与传统偏最小二乘回归(PLS)及支持向量机(SVM)技术加以系统比较.结果表明,GP建模性能不亚于广泛使用的SVM,显著优于经典PLS.此外,GP能够较好处理线性和非线性混合问题、自动确定模型结构、能够通过超参数解释体系噪音纳入和变量贡献,给出预测结果的置信评估,这些特点皆是传统方法所不具备的.鉴于此,可以认为GP是一种具有开发潜力的机器学习策略,不仅可供分析域/肽识别和相互作用,还可用于解决和处理其他生物相关问题.更多还原

【Abstract】 Many protein-protein interactions involved in cell signaling networks conduct with the manner so-called folding-on-binding,which are mediated by the binding of a globular domain in one protein to a short peptide stretch in another.Thus,systematic analysis and reliable prediction of domain-peptide recognition and interaction are fundamentally important for our understanding of the molecular mechanism and biological implications underlying cell signaling.Herein,we report the use of a new and powerful machine learning technique called Gaussian process(GP) to carry out statistical modeling and structural analysis for four categories of domain-peptide systems,including SH3,PDZ,14-3-3,and GYF domains.The results of the modeling are compared systematically to those deriving from classical partial least square(PLS) regression and sophisticated supporting vector machine(SVM).We demonstrate that GP is comparable with or even better than nonlinear SVM,and is much well to linear PLS.In addition,GP possesses some more merits as it is capable of handling linearity and nonlinearity-hybrid problems effectively,determining algorithmic parameters automatically,interpreting obtained models straightforwardly,and providing additional evaluation for predictions quantitatively.All of these come together to suggest that GP would be a promising tool not only for exploring the domain-peptide interaction behavior,but also for solving other chemistry and biology-related problems.更多还原