基于深度学习的蛋白质残基相互作用预测

发布时间：2018-01-13 08:34

本文关键词：基于深度学习的蛋白质残基相互作用预测　出处：《苏州大学》2016年硕士论文　论文类型：学位论文

【摘要】：残基对的相互作用描述了蛋白质三维结构中一对残基的空间距离关系,相互作用的残基对对维护蛋白质结构的稳定起着重要作用。蛋白质中所有残基对的相互作用关系确定了蛋白质三维结构的二维拓扑,所以得到准确的残基对相互作用关系对蛋白质三维结构预测有重要意义。残基对相互作用预测,特别是长范围残基对相互作用预测的准确率一直很低。这主要是因为残基对特征与残基对相互作用的高度非线性,另外残基对正负样本比例的严重失调也降低了模型的泛化能力。本文研究了基于双向递归神经网络的深度序列模型以及减轻样本比例失调影响的训练算法。双向递归神经网络模型不仅可以接收变长的蛋白质序列特征,而且它在处理残基特征的时候也不需要指定滑动窗口大小,但滑动窗口却是普通浅学习方法所需要的。本文的训练算法在控制正负样本比例的同时,动态地选择输入给分类模型的样本。深度神经网络通过大量非线性变换把原始特征转换为高级特征,这种变换很适合残基相互作用预测这样的应用问题,但深度神经网络由于包含多层神经网络又会使超参数的选择成为难题。本文基于Hyperopt超参数优化框架实现了深度序列模型的并行超参数优化。通过快速的并行搜索,我们找到了一个与人工花费大量时间搜索到的模型不相上下的模型。这个模型在多个测试集上获得的中范围残基相互作用预测准确率超过其它方法10%以上,在长范围残基相互作用上的预测结果和当前流行方法不相上下。
[Abstract]:The interaction of residue pairs describes the spatial distance relationship of a pair of residues in a protein three-dimensional structure. The interacting residues play an important role in maintaining the stability of protein structure. Therefore, it is very important to obtain accurate residue pair interaction relationship for protein three-dimensional structure prediction and residue pair interaction prediction. Especially, the prediction accuracy of long-range residue pair interaction is very low, which is mainly due to the highly nonlinear interaction between residue pair characteristics and residue pair. In addition, the serious imbalance of residue to the positive and negative sample ratio also reduces the generalization ability of the model. In this paper, the depth sequence model based on bidirectional recurrent neural network and the training algorithm to reduce the effect of sample misalignment are studied. The recurrent neural network model can not only receive variable length protein sequence features. And it does not need to specify the size of sliding window when dealing with residual feature, but the sliding window is required by ordinary shallow learning methods. The training algorithm in this paper is to control the positive and negative sample ratio at the same time. The deep neural network transforms the original features into high-level features through a large number of nonlinear transformations, which is suitable for the application of prediction of residual interaction. However, the selection of superparameters in depth neural networks will become a difficult problem due to the inclusion of multi-layer neural networks. In this paper, the parallel hyperparametric optimization of depth sequence models is realized based on Hyperopt hyperparametric optimization framework. Fast parallel search. We have found a model which is comparable to the model which takes a lot of time to search. The prediction accuracy of mid-range residue interaction on multiple test sets is more than 10%. Predictions on long range residue interactions are comparable to current popular methods.
【学位授予单位】：苏州大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：Q51;TP181

【相似文献】