Hierarchical Representation and Graph Convolutional Networks for the Prediction of Protein–Protein Interaction Sites

Proteins carry out a broad range of functions in living organisms usually by interacting with other molecules. Protein–protein interaction (PPI) is an important base for understanding disease mechanisms and for deciphering rational drug design. The identification of protein interactions using experimental methods is expensive and time-consuming. Therefore, efficient computational methods to predict PPIs are of great value to biologists. This work focuses on predicting protein interfaces and investigates the effect of different molecular representations in the prediction of such sites. We introduce a molecular representation according to its hierarchical structure. Therefore, proteins are abstracted in terms of spatial and sequential neighboring among amino acid pairs, while we use a deep learning framework, Graph Convolutional Networks, for data training. We tested the framework on two classes of proteins, Antibody–Antigen and Antigen–Bound Antibody, extracted from the Protein–Protein Docking Benchmark 5.0. The obtained results in terms of the area under the ROC curve (AU-ROC) on these classes are remarkable.