基于图注意力网络和简单循环单元的化合物-蛋白质交互预测

Compound-Protein Interaction Prediction Based on Graph Attention Network and Simple Recurrent Unit

现有基于深度学习的化合物-蛋白质交互预测方法未考虑数据的内部协变量偏移及序列数据的长距离依赖.针对此问题,文中提出基于图注意力网络和简单循环单元的化合物-蛋白质交互预测方法.利用图注意力网络-门控循环单元学习化合物分子的图级表示,利用多层简单循环单元学习氨基酸子序列的特征向量表示,结合多层前馈网络预测化合物-蛋白质的交互作用.实验表明,文中方法在2个公开数据集上的各项评估指标都有所提升,由此验证方法的有效性.

The internal covariant shift of the data and the long distance dependence of the sequence data are not taken into account in the existing deep learning based compound protein interaction prediction methods.To solve the problem,a method for compound-protein interaction prediction based on graph attention network and simple recurrent unit is proposed.The graph attention network-gated recurrent unit is introduced to learn the graph-level representation of compound molecules,the multi-layer-simple recurrent unit is employed to learn feature vector representation of amino acid subsequences,and multilayer-feed-forward network is utilized to predict compound-protein interactions.Experiments show that the evaluation indexes of the proposed method are improved on 2 public datasets,and the effectiveness of the proposed method is verified.