Graph Neural Network for the prediction of Antibiotic Resistance.

Antimicrobial resistance is one of the main threats to global health. Antibiotics still represent the most reliable solution for bacterial infection treatment, but the spread of antibiotic resistance threatens their efficacy resulting in a dramatic worldwide increase in morbidity and mortality. Each bacterial species is often only susceptible to a few specific antibiotics; this is why it is critical to identify the correct and still effective antibiotic to be used in a clinical setting. Although genetic testing of bacteria is increasingly used in the medical lab, a time-consuming antibiogram, based on bacterial cultures, is still the standard approach. Current techniques make it possible to sequence bacterial DNA in a much shorter time, allowing the discrimination of a resistant bacterium from a susceptible one based on the presence of specific elements that confer resistance. This thesis aims to perform this task on general basis, classifying bacteria genomes through Graph Neural Networks (GNN). A GNN represents a particular deep learning technique developed first time in 2009 as a promising solution for dealing with graph-structured data. The raw data that constitute the bacterial genome can be transformed into particular graphs, such as De Bruijn Graph, and then used to train a GNN. In this context, the primary methods are described and compared, focusing on how the convolution in Euclidean space could be extended to the case of graphs. For the genome of Staphylococcus aureus, in the case of several antibiotics, such as Oxacillin, Erythromycin, Ciprofloxacin, the experimental evaluation demonstrates that Graph Sage convolution, a particular operator suitable for processing large graphs, has been shown to classify the characteristic resistance/susceptibility correctly. The mean accuracy obtained with stratified k-fold cross-validation is higher than the 80% for Ciprofloxacin. This study, clearly, demonstrates the possibility of predicting the occurrence of the phenomenon of antimicrobial resistance with GNN models. It can therefore aspire to help in evaluations and clinical studies.