Exploring neural networks to uncover information-richer features for protein interaction prediction.

Moving in a crowded cellular environment, proteins have to recognize and bind to each other with high specificity. This specificity reflects in a combination of geometric and chemical complementarities at the core of interacting regions that ultimately influences binding stability. Exploiting such peculiar complementarity patterns, we recently developed CIRNet, a neural network architecture capable of identifying pairs of protein core interacting residues and assisting docking algorithms by rescaling the proposed poses. Here, we present a detailed analysis of the geometric and chemical descriptors utilized by CIRNet, investigating its decision-making process to gain deeper insights into the interactions governing protein-protein binding and their interdependence. Specifically, we quantitatively assess (i) the relative importance of chemical and physical features in network training and (ii) their interplay at protein interfaces. We show that shape and hydrophobic-hydrophilic complementarities contain the most predictive information about the classification outcome. Electrostatic complementarity alone does not achieve high classification accuracy but is required to boost learning. Ultimately, our findings suggest that identifying the most information-dense features may enhance our understanding of the mechanisms driving protein-protein interactions at core interfaces.