Neural network–based approach for improving the evaluation of antibody–antigen docking poses

The role of artificial intelligence (AI)–based approaches in computational biology and molecular biophysics has become increasingly central over the past decade; however, many challenges remain unresolved, such as the accurate prediction of protein–protein complexes, the complete solution of which would have a significant impact both on our understanding of cellular mechanisms and on the development of therapeutic and diagnostic strategies. Here, we present a protocol based on multiple minimal neural network (NN)–based approaches, trained on a set of carefully selected physicochemical features, to discriminate docking decoy poses (structurally distant from the experimental complex) from native-like poses (structurally close to the native conformation) within a specific class of biologically relevant protein–protein complexes, namely antibody–antigen systems in which the antigen is a protein. A specific version of the proposed method, trained on a set of antibody–antigen interface descriptors, some of which are derived from graph theory to capture the geometric complexity of intermolecular interactions, was compared with ITScore-PP, the docking score provided by HDOCK. This NN-based approach, demonstrates the ability not only to distinguish native-like poses from decoys, but also, more challengly, to discriminate intermediate poses from native-like ones. Furthermore, it was also able to predict the DockQ score, a widely used metric for assessing docking pose quality, showing a larger absolute Pearson correlation coefficient than ITScore-PP. The ability of our NN-based approach, which relies solely on structural interface features, to identify accurate dockings highlights its potential as a valuable tool for improving the ranking of antibody–antigen docking poses and underscores the importance of sppropriate feature selection in protein-protein interaction modeling.