Real-SPINE: An integrated system of neural networks for real-value prediction of protein structural properties
Real-SPINE: An integrated system of neural networks for real-value prediction of protein structural properties
复制标题
DOI：
10.1002/prot.21408
复制
发表时间：
2007-07-01
影响因子：
2.9
通讯作者：
Zhou, Yaoqi
中科院分区：
文献类型：
--
作者：
Dor, Ofer;Zhou, Yaoqi
蛋白质可以在三维空间中自由移动。因此，它们的结构特性，例如溶剂可及表面积、主链二面角和原子距离，是连续变量。然而，这些属性通常被任意分为几类，以便于通过统计学习技术进行预测。在这项工作中，我们建立了一个用于实值预测的神经网络集成系统（称为 Real-SPINE），并应用该方法仅基于序列衍生的信息来预测蛋白质的残留溶剂可及性和主链 Psi 二面角。 Real-SPINE 使用包含 2640 条蛋白质链的大型数据集、多重序列比对生成的序列图谱、代表性氨基酸特性、缓慢的学习速率、过度拟合保护和预测的二级结构进行训练。该方法优化了超过 200,000 个权重，并在预测和实际溶剂可及表面积之间产生 10 倍交叉验证的皮尔逊相关系数 (PCC)，为 0.74，在预测和实际 Psi 角度之间为 0.62。特别是，2640 种蛋白质中 90% 的预测溶剂可及表面积与实际溶剂可及表面积之间的 PCC 值大于 0.6。 Real-SPINE 的结果可以与最佳报告的溶剂可及表面积相关系数 0.64-0.67 和 Psi 角相关系数 0.47 进行比较。 real-SPINE 服务器、可执行程序和数据集可在 http://sparks.informatics.iupui.edu 上免费获得。
Proteins can move freely in three-dimensional space. As a result, their structural properties, such as solvent accessible surface area, backbone dihedral angles, and atomic distances, are continuous variables. However, these properties are often arbitrarily divided into a few classes to facilitate prediction by statistical learning techniques. In this work, we establish an integrated system of neural networks (called Real-SPINE) for real-value prediction and apply the method to predict residue-solvent accessibility and backbone Psi dihedral angles of proteins based on information derived from sequences only. Real-SPINE is trained with a large data set of 2640 protein chains, sequence profiles generated from multiple sequence alignment, representative amino-acid properties, a slow learning rate, overfitting protection, and predicted secondary structures. The method optimizes more than 200,000 weights and yields a 10-fold cross-validated Pearson's correlation coefficient (PCC) of 0.74 between predicted and actual solvent accessible surface areas and 0.62 between predicted and actual Psi angles. In particular, 90% of 2640 proteins have a PCC value greater than 0.6 between predicted and actual solvent-accessible surface areas. The results of Real-SPINE can be compared with the best reported correlation coefficients of 0.64-0.67 for solvent-accessible surface areas and 0.47 for Psi angles. The real-SPINE server, executable programs, and datasets are freely available on http://sparks.informatics.iupui.edu.