Combining deep graph convolutional networks and PRSA to enhance protein-protein interaction site prediction

Abstract

Protein-protein interaction(PPI) site prediction is a deep-level exploration of the mechanism of life activity, but relying solely on experimental methods to identify PPI sites is hugely costly. This method is advantageous among the developed computational methods using structural information. For the relative solvent accessibility (RSA) of protein structural information, the absolute values of solvent accessibility derived from the program named DSSP (Kabsch and Sander, 1983) were primarily used and then normalized using the highest exposure area of the amino acid type determined in the past. It is difficult to obtain suitable RSA when protein structure information cannot be obtained by homologous transfer, and thus the use of RSA is limited. We used the latest deep learning prediction tools to mine potentially valuable information from long-range interactions inside protein sequences and used it for protein RSA prediction. In a deep graph convolutional neural network, we incorporate the predicted relative solvent accessibility (PRSA) into the original structural information and then combine the sequence information and evolutionary information to form graph node features. We showed that our proposed method significantly improves the performance of AUPRC and MCC by over 9.5% and 21% compared to other sequence-based and structure-based methods. Furthermore, it was demonstrated by analyzing the method that the PRSA plays a crucial role in PPI site prediction.