利用BP神经网络对蛋白质折叠速率的研究

Prediction of protein folding rates by the BP neural network model

目的:通过BP神经网络模型对蛋白质的折叠速率进行预测。方法:以蛋白质的接触序参数(CO、LRO、TCD)为输入信息,折叠速率为输出信息,构建具有1个输入层、2个隐藏层和一个输出层的4层BP神经网络模型,用以尝试多种激活函数,给出合理的预测结果。结果:激活函数是影响BP神经网络效率和精度的最重要因素之一,文中比较了6个激活函数来评估神经网络模型在蛋白质折叠预测中的表现,发现Mish函数预测全β蛋白质比较有效,而Tanh函数更适合一般性蛋白质折叠速率的预测。本文还通过不同的分离截止值来了解蛋白质的序列和折叠率之间的关系,LRO和TCD/CO的分离截止值分别为8和3的模型可以更精确地预测折叠率。结论:在本模型中,蛋白质折叠速率预测结果和实验结果基本吻合,这些研究成果为寻找影响蛋白质折叠过程的动力学因素提供了理论依据。

Aims:This paper aims to study the prediction of the folding rates of proteins by the BP neural network model.Methods:The contact parameters(CO,LRO and TCD)were set as input nodes.The output node was folding rate.A 4-layer neural network was constructed;and six activation functions were chosen to provide effective prediction results.Results:The results showed that the activation function Tanh in the BP neural network had higher prediction accuracy to predict general proteins while Mish was suitable forβ-protein.It was found that the model with the separation cutoff values of LRO and TCD/CO at 8 and 3 respectively could be used to predict the folding rate more precisely.Conclusions:The model achieved good agreement between prediction and experimentally observed folding rates of proteins.The result provides a theoretical basis for better understanding the key factor of the protein folding process.