Objectives Protein-protein interaction (PPI) is an important part of many life activities in organisms. Although a large number of PPIs have been verified by high-throughput techniques in the past decades, currently known PPI pairs are still far from complete.

Methods In order to improve the feature extraction methods of prediction performance, we used conjoint triad (CT), auto covariance (AC), local descriptor (LD) and AC+CT, four kinds of feature extraction methods to build DNN models based on deep neural networks.

Results The results showed that the model DNN-CT achieved superior performance with accuracy of 97.65%, recall of 98.96%, area under the curve (AUC) of 98.51% and loss of 26.69%, respectively. Although the performance of the DNN-LD was not prominent, the trends of various indicators were relatively stable, and achieved an accuracy of 95.30%, recall of 98.28%, AUC of 95.57% and loss of 36.23%, respectively.

Conclusions By comparison, we found that DNN-CT and DNN-LD were superior to DNN-AC and DNN-(CT+AC). The results of our experiment can provide a supplementary tool for future proteomics study.