Impact Award
Team 13344: Chiara Mastrogiuseppe (Team lead, Humanitas University), Luca Busani (Istituto Superiore di Sanità), Andrea Cacioppo (Istituto Superiore di Sanità), Francesco Cellitti (Sapienza University of Rome), and Guido Gigante (Istituto Superiore di Sanità)
Antimicrobial resistance surveillance is usually observed through pathogen–antibiotic susceptibility testing: for each pair, laboratories report susceptible, intermediate, and resistant isolate counts as phenotypic traits of the bacterium. However, these S/I/R measurements are sparse and noisy observations of a deeper process: the evolution of an underlying AMR landscape across pathogens, antibiotics, places, and time. This project aims to model that latent AMR landscape and use it to reconstruct, compare, and generate uncertainty-aware short-term projections of changes in resistance profiles.
We propose to represent each surveillance snapshot as a partially observed pathogen– antibiotic graph. Observed edges contain S/I/R counts or proportions, while missing edges are treated as unobserved rather than as biologically negative observations. Using Pfizer ATLAS as the primary data source, and additional AMR Register datasets where harmonization is feasible, we will aggregate isolate-level records into country–year pathogen–antibiotic S/I/R count matrices. A deep probabilistic framework will be used to infer a single posterior distribution over latent AMR states, from which matrix completion, short-term forecasting, and anomaly detection can be derived.
For each country–year surveillance unit, an encoder will map the partially observed pathogen–antibiotic susceptibility matrix into a low-dimensional latent AMR state. A decoder will then reconstruct calibrated S/I/R distributions for both observed and unobserved pairs, using an overdispersed likelihood such as a Dirichlet-multinomial where appropriate. Temporal evolution of the AMR landscape will be modeled probabilistically in this low-dimensional latent space, enabling short-term projections through sampled latent trajectories with associated uncertainty. Prior knowledge can be incorporated through pathogen and antibiotic embeddings, including taxonomic or resistance-mechanism descriptors where available.
The resulting model will support completion of sparse pathogen–antibiotic matrices, uncertainty-aware short-term projections, and anomaly detection through unexpectedly low predictive-likelihood observations. Since S/I/R profiles are shaped by both biological resistance evolution and surveillance practice — including testing coverage, breakpoint definitions, and sampling composition — the model will serve as a surveillance-support framework to make incomplete AMR data more informative for monitoring, comparison, and prioritisation of pathogen–drug combinations with worsening or uncertain inferred profiles.