背景:溶剂可及性的预测可以为分析蛋白质的结构和功能提供有价值的线索,如蛋白质3维结构和B细胞表位预测。为了完全破译蛋白质-蛋白质相互作用的过程,最初但关键的一步是计算蛋白质溶剂的可及性,特别是当蛋白质的三级结构未知时。尽管已经在蛋白质溶剂可及性预测方面做出了一些努力,现有方法的性能远远不能令人满意。
方法:为了开发高精度模型,我们关注一些关于预测性能的可能方面,包括几个序列派生的特征,加权滑动窗口方案和机器学习方法的参数优化。为了解决上述问题,我们采取以下策略。首先,我们探索了已观察到与残留溶剂可及性相关的各种特征。这些区别特征包括蛋白质进化信息,预测蛋白质二级结构,原生病症,物理化学倾向和几个基于序列的残基结构描述符。其次,观察到滑动窗口中相邻残基的不同贡献,因此,提出了一种加权滑动窗口方案来区分相邻残基对中心残基的贡献。第三,粒子群优化(PSO)用于搜索所提出的预测器的全局最佳参数。
结果:通过3倍交叉验证评估,对于我们新编制的数据集,我们的方法实现了14.1%的平均绝对误差(MAE)和0.75的人相关系数(PCC).与两个基准数据集中最先进的预测模型相比,我们的方法表现出更好的性能。实验结果表明,我们的PSAP实现了高性能,优于许多现有的预测因子。名为PSAP的Web服务器已构建,可在http://59.73.198.144:8088/SolventAccessibility/免费获得。
BACKGROUND: The prediction of solvent accessibility could provide valuable clues for analyzing protein structure and functions, such as protein 3-Dimensional structure and B-cell epitope prediction. To fully decipher the protein-protein interaction process, an initial but crucial step is to calculate the protein solvent accessibility, especially when the tertiary structure of the protein is unknown. Although some efforts have been put into the protein solvent accessibility prediction, the performance of existing methods is far from satisfaction.
METHODS: In order to develop the high-accuracy model, we focus on some possible aspects concerning the prediction performance, including several sequence-derived features, a weighted sliding window scheme and the parameters optimization of machine learning approach. To address above issues, we take following strategies. Firstly, we explore various features which have been observed to be associated with the residue solvent accessibility. These discriminative features include protein evolutionary information, predicted protein secondary structure, native disorder, physicochemical propensities and several sequence-based structural descriptors of residues. Secondly, the different contributions of adjacent residues in sliding window are observed, thus a weighted sliding window scheme is proposed to differentiate the contributions of adjacent residues on the central residue. Thirdly, particle swarm optimization (PSO) is employed to search the global best parameters for the proposed predictor.
RESULTS: Evaluated by 3-fold cross-validation, our method achieves the mean absolute error (MAE) of 14.1% and the person correlation coefficient (PCC) of 0.75 for our new-compiled dataset. When compared with the state-of-the-art prediction models in the two benchmark datasets, our method demonstrates better performance. Experimental results demonstrate that our PSAP achieves high performances and outperforms many existing predictors. A web server called PSAP is built and freely available at http://59.73.198.144:8088/SolventAccessibility/.