Title:Compact assessment of molecular surface complementarities enhances neural network-aided prediction of key binding residues

Abstract:Predicting interactions between biomolecules, such as protein-protein complexes, remains a challenging problem. Despite the many advancements done so far, the performances of docking protocols are deeply dependent on their capability of identify binding regions. In this context, we present a novel approach that builds upon our previous works modeling protein surface patches via sets of orthogonal polynomials to identify regions of high shape/electrostatic complementarity. By incorporating another key binding property, such as the balance between hydrophilic and hydrophobic contributions, we define new binding matrices that serve an effective inputs for training a neural network. Our approach also allows for the quantitative definition of a typical binding site area - approximately 10Å~in radius - where hydrophobic contribution and shape complementarity, which reflects the Lennard-Jones interaction, are maximized. Using this new architecture, CIRNet (Core Interacting Residues Network), we achieve an accuracy of approximately 0.82 in identifying pairs of core interacting residues on a balanced dataset. In a blind search for core interacting residues, CIRNet distinguishes these from decoys with a ROC AUC of 0.72. This protocol can enahnce docking algorithms by rescaling the proposed poses. When applied to the top ten models from three popular docking server, CIRNet improves docking outcomes, reducing the the average RMSD between the refined poses and the native state by up to 58%.