Prediction of protein interactions based on CT-DNN

Protein-protein interactions (PPIs) play critical roles in many cellular biological processes, underlie the entire molecular machinery of living cells, and can be used to aid in drug target detection and therapeutic design. Traditional methods such as biochemical analysis and chromatography have low efficiency and low coverage for PPI identification. In recent years, the amino acid sequence information of proteins has developed rapidly, and machine learning methods based on amino acid sequences have been widely developed. It is difficult to improve the PPI prediction accuracy of traditional machine learning, and it is difficult to deal with the increasing protein sequence data, which hinders the application of traditional machine learning in protein interaction prediction. Deep learning methods simulate the learning mechanism of the human brain and have been successfully applied in speech recognition and image recognition, natural language understanding and other fields. Compared with traditional machine learning methods, deep learning algorithms can handle large-scale raw and complex data and can automatically learn useful abstract features. Deep neural networks (DNNs) have brought a turnaround in protein interaction prediction, which can process large-scale protein sequence data and automatically extract low-level or high-level features required for classification. In this paper, we propose a CT-DNN method to predict PPI, which takes the protein sequence as the research basis, considers the properties of an amino acid and its neighboring amino acids, and divides the amino acids into 7 amino acids according to the dipolar polarity and volume of the amino side chain category, take three consecutive amino acids as a unit to obtain the projection vector space of the protein sequence, and normalize it. By constructing protein interaction relationship pairs, a deep neural network (DNN) method is introduced for learning. By adjusting the learning rate, Different parameters such as activation function are used to improve the prediction performance of the model. The best model is trained through 10-fold cross-validation, on the test set, auc is 0.983 and aupr is 0.982. Compared with other commonly used models, it is proved that CT-DNN has has good learning and generalization ability, and can be effectively used for PPI prediction.