GCNSP: A Novel Prediction Method of Self-Interacting Proteins Based on Graph Convolutional Networks

As an essential protein interaction, self-interacting proteins (SIPs) plays a vital role in biological processes. Identifying and confirming SIPs is of great significance for the exploration of new gene functions, protein function research and proteomics research. Although a large number of SIPs have been confirmed with the rapid development of high-throughput technology, the biological experimental method is still limited by blindness and high cost, and has a high false-positive rate. Therefore, the use of computational techniques to accurately and efficiently predict SIPs has become an urgent need. In this study, a novel SIPs prediction method GCNSP based on Graph Convolutional Networks (GCN) is proposed. Firstly, the evolution information of protein is described by Position-Specific Scoring Matrix (PSSM). Then the feature information is extracted by GCN, and finally fed into Random Forest (RF) classifier for accurate classification. In the five-fold cross-validation on Human and Yeast data sets, GCNSP achieved 93.65% and 90.69% prediction accuracy with 99.64% and 99.08% specificity, respectively. In comparison with different classifier models and other existing methods, GCNSP shows strong competitiveness. The excellent results show that the proposed method is very suitable for SIPs prediction and can provide highly reliable candidates for biological experiments.