Prediction of Protein-ATP Binding Sites Based on Word Vector Convolution Model

Abstract

Recent studies have shown that the interaction between protein and ATP is closely related to human diseases, and the ATP-binding sites in protein sequences have become the focus of drug design. In order to improve the prediction accuracy of Protein-ATP binding sites, in this paper, we propose a prediction method based on word vector convolution neural network. Firstly, we extract five types of features from protein sequences including the position specific scoring matrix, protein secondary structure, solvent accessible surface area, sequence characteristics and residue physicochemical property. Then, the RepeatedEditedNearestNeighbours method is used to clean the data, and the sample imbalance problem is solved by random under-sampling. The under-sampled data is encoded by word vectors. Finally, the improved deep convolution neural network model is trained and compared with the related prediction methods. The experimental results show that our proposed prediction method can predict the Protein-ATP binding sites more precisely.