Unveiling disease-protein associations by navigating a structural alphabet-encoded protein network

The identification of genes associated with human disorders is a major goal in computational biology. Although the rapid emergence of cellular network-based approaches has been successful in many instances, all of these methodologies are partially limited by the incompleteness of the interactome. Here, we propose a novel method that may overcome the inherent problem of these incomplete molecular networks and assist already established network techniques. Instead of using protein-protein interaction networks, we encode the local threedimensional structure of a protein into a series of letters, called the Structural Alphabet, and define a proteomic structural network in which each node represents a unit of the structural alphabet (USA) and each pair of USAs is linked based on their structural similarity. This novel structural network is the platform by which a diffusion-based algorithm determines the potential involvement of proteins in disease phenotypes. Computational experiments show that the combination of diffusion-based methods with the constructed structural alphabet network offers better predictive performance than the results obtained using interactome networks and provides a new avenue to assist in identifying disease-related proteins.