Traditional Network Analysis and Graph Neural Networks Help in Prioritizing Repurposable Drugs to Treat Multiple Sclerosis

Poster for the 23rd European Symposium on Quantitative Structure-Activity Relationship (Heidelberg, Germany 2022) Abstract Drug repurposing consists in identifying additional uses for known drugs and, since these new findings are built on previous knowledge, it reduces both the length and the costs of the drug development. In this work, we want to integrate traditional network analysis techniques with graph neural network methods for prioritizing plausible drugs for repurposing to treat multiple sclerosis (MS). MS is both the most frequent non-traumatic disabling disease in young adults and the commonest demyelinating disease. The etiology and the mechanism causing its worsening progression are still unclear, nevertheless it has been proven that a complex interplay of genetic and environmental factors is important. In recent years, a plethora of computational approaches to drug repurposing have been proposed and applied to a wide variety of therapeutic areas. Most of such approaches rely either on machine learning or on the traditional methods of computational drug design. Network-based models, as the ones we utilize in this study, could be adequate theoretical tools for investigating multifactorial disorders as multiple sclerosis. They would allow us to take into account the latent complex structure of this disease without losing a comprehensive view. One of such approaches takes the human protein-protein interactome as a reference network to quantify the relatedness between drugs and diseases by calculating the distance between drug targets and disease-associated proteins. This distance has been proposed as a suitable metrics to measure the "proximity" between drugs and diseases. Here, we want to support traditional network analysis with deep graph neural representations in order to consider also more complex interactions. Graph neural networks (GNN) are deep neural networks specialized on graph-structured data, they learn a vector representation of nodes (vertices, entities), edges (links, connections) or even entire graphs. This representation (embedding) collects both the information coming from the local neighborhood of a node as well as the one from the global topological structure of the graph, facilitating the identification of multimodal and complex relationships. Combined therapies are then prioritized on the basis of the drugs’ separation on the human interactome and the known drug-drug interactions. Desired combinations are those ones mixing drugs being topologically and pharmacologically distinct. In this case, the two drugs synergistically cooperate in treating the disease, yet each one in its own way. Employing traditional network analysis, we have already collected a number of drugs and their plausible combinations that could be proposed for the treatment of multiple sclerosis.