Finalist
Team 13355: Giordano De Marzo (Team lead, Konstanz University), Alfredo De Bellis (Fondazione Bruno Kessler, Center for Health Emergencies), and Chiara Mastrogiuseppe (Humanitas University)
Antimicrobial resistance surveillance is essential for guiding empirical treatment, antimicrobial stewardship, and public-health prioritisation. Yet most surveillance outputs are based on pairwise susceptibility indicators, such as the proportion of isolates classified as susceptible or resistant for each pathogen–antibiotic combination. These measures are essential, but they treat each combination in isolation, whereas clinical and policy decisions depend on the structure of treatment options as a whole. In fact, susceptibility records naturally form a bipartite network in which pathogens are linked to the antibiotics that remain effective against them, a structure that contains information on redundancy, fragility, and dependence on last-line or broadly active antibiotics that is invisible when each pair is analysed separately.
This project will use Pfizer ATLAS surveillance dataset to build and analyse the pathogen– antibiotic bipartite networks. We will characterise their topology, including modularity, nestedness, overlap in antibiotic coverage, and uniqueness of treatment options, and use bipartite projections to derive interpretable similarity measures among pathogens and antibiotics. Building on this representation, we will adapt the Fitness–Complexity (FC) algorithm, originally developed in economic complexity, to rank antibiotics and pathogens from the structure of effective coverage.
Within this framework, an antibiotic has high fitness when it provides broad and structurally important coverage across the network, especially for pathogens with few alternatives. A pathogen has high complexity when its remaining effective options depend on antibiotics that are themselves central to the network. This matters because the value of an antibiotic is not captured by its average activity alone: it depends on which pathogens it covers, how many alternative active agents those pathogens have, and whether its coverage is shared by other effective antibiotics or occupies a unique position in the network.
Crucially, network-based methods of this kind can recover the predictive accuracy of machine learning models while remaining more interpretable, a critical advantage in clinical and policy contexts.
The public-health value of this approach is that it moves beyond describing resistance percentages toward identifying structural warning signals in the treatment landscape. We will validate the framework by assessing whether network-derived indicators recover known clinical and public-health priorities, identify low-coverage or difficult-to-treat pathogen profiles and add information beyond conventional pairwise resistance summaries. These outputs can support antimicrobial stewardship, awareness of empirical treatment vulnerabilities, and surveillance prioritisation by highlighting where loss of activity would have the greatest system-level consequences. The project is supported by a young multidisciplinary team combining expertise in network science, biomedical AI/ML, and infectious-disease modelling.