Tackling antimicrobial resistance in hospitals: a holistic eco-evolutionary model of resistance gene dissemination to optimize intervention strategies – ResisTrack

ResisTrack relies on the following approaches:
1. An original method of multiscale ecosystem simulation based on the representation of nested elements, such as genes, plasmids, bacteria or patients, as an inclusion graph. The method behaves as a hybrid between individual-based models, with maximal flexibility but poor computational efficiency, and population based models that can simulate large populations (and capture rare events) but with poorer flexibility.
2. A systematic mapping of the link between the bacterial resistome (the genomic content of ARGs and mobile genetic elements harboring ARGs) and routine-generated susceptibility profile. In turn, this mapping is employed to infer the resistome of all bacteria diagnosed at a large hospital group, taking inference uncertainty into account.
3. A calibration of the ecosystem model parameters based on real hospital data reflecting inferred resistomes, antibiotic treatments and patient movements in the hospital network.
WP1 implements a computationally efficient, stochastic, generative model of a hospital ecosystem called msevol, and will determine priors for parameters to be inferred in WP3. WP2 objective is to build a comprehensive dataset describing ARGs and MGEs in enterobacteria, microbiological diagnoses and antibiotic prescriptions. In WP3, we will calibrate the stochastic msevol model by learning parameters from hospital data. In WP4, we will implement, evaluate and select candidate AS/IC strategies by monitoring each strategy outcome in a simulated hospital ecosystem.
Patient-level parameters are estimated from available hospital data, including the rate of transfer between wards, the probability of receiving antibiotics when in a given ward, and the probability of being sampled for microbiological tests, for either screening or diagnosis. The estimation of microbiological parameters, such as conjugation rates, fitness cost and transmission rates, cannot be performed directly and will rely on approximate Bayesian computation with carefully chosen parameter priors. Our eco-evolutionary analyses will focus on enterobacteria, in which HGT is most relevant to resistance. To estimate the distribution of ARGs and MGEs in all hospital enterobacteria, we need first to determine which isolate harbors which ARG or MGE. We will infer the resistomes from routine phenotypic findings, using a probabilistic model trained on the genomes of a reduced set (n ~ 1,600) of isolates representing the diversity of AST profiles found in the HCL, augmented with other data sources.

In WP1, a functional version of msevol implements antibiotic pressure, ARGs, gene transfer with plasmids, bacterial diffusion between niches and competition for resources between bacterial species. Using msevol, we examined the conditions of the persistence of plasmid-borne bacterial multi-resistance, mimicking the competition between high-concern resistance plasmids in enterobacteria, pOxa48, pCTX-M15 and pNDM-1. Simulation results have identified the conditions that favor the accumulation of several plasmids (pCTX-M15 and pOxa48) to confer multi-drug resistance, as well as the conditions that favor the persistence of a single plasmid with stronger fitness cost (pNDM-1). The current version will be augmented with patient movements between wards, antimicrobial treatment initiation and duration, intragenomic plasticity (transposon movement within the same bacterial genome). In parallel, a deterministic model of intragenomic ARG mobility was implemented to study the emergence of co-transferable ARG clusters involved in the dissemination of multidrug resistance. Based on this model, Partner 1 identified potential driving forces of ARG clustering such as the coexistence of MDR and susceptible species at the interface between environments with high and low antibiotic pressure.

Model parameters have been defined and estimated from data of the literature, using a meta-analysis approach. Genome plasticity parameters have been defined by Partner 3 as a function of the pan-genome size evolution relative to the number of sampled genomes in a species. Patient movements between wards have been collected. Parameters regarding between-patient transmission of bacteria under MDR/XDR contact precautions have been discussed with infection control experts and will rely on literature search (ongoing).

In WP2, a collection of reference strains for National Reference Centers has been consolidated. This collection contains more than 300 strains of P. aeruginosa, ~240 of E. coli, ~170 of A. baumanii, ~140 of K. pneumonia, and other enterobacteria species. Whole-genome sequencing using both Nanopore and Illumina technologies has been implemented with a throughput of ~50 genomes/week. A prototype resistome prediction model has been constructed to study the prediction of ARGs (but not yet plasmids) from resistance profiles.

In WP3, we identified state-of-the art methods to calibrate the stochastic msevol model by learning parameters from hospital data. Two R packages synlik (Fasiolo and Wood, 2021) and BSL (An et al, 2022) implement the synthetic likehood method developed by Wood (2010) and incorporate MCMC to find the approximative posterior distribution. These methods are currently compared to identify the most efficient approach.

Candidate strategies against AMR. By constructing and training a realistic simulation model of a bacterial hospital ecosystem, ResisTrack will allow to evaluate an exhaustive set of strategies against antibiotic resistance and to monitor their epidemiological outcome and cost-benefit balance. These results will provide a set of well-defined candidate optimal strategies, along with ecological explanations of their impact on the dissemination of ARGs.

A general model of bacterial resistance in the hospital. Our ARG-centric ecosystem model will be adaptable to other hospitals and settings, thanks to the clear separation between global, setting-independent parameters, and local parameters. ResisTrack will provide a general simulation framework to tailor interventions against AMR in the light of the local epidemiology.

A resistome inference model. The analysis of an exhaustive set of bacterial genomes along with phenotypic resistance data will provide an inference model able to determine the resistome (ARG and MGE content) of an isolate from standard-of-care microbiological data. Beyond ResisTrack, this approach will pave the way for breakthrough applications in clinical microbiology, including the real-time monitoring of the epidemiology of ARGs or the correction of AST results based on their consistency with a plausible set of ARGs. The model’s application will allow hospitals to monitor the dissemination of ARGs and plasmids at no added cost.

Findings regarding the driving forces of ARG clustering and co-transfer have been presented at the EvoLyon conference in 2021.
Partner 1 will present results at the ECML/PKDD 2022 conference (MLMG 2022: workshop on machine learning for microbial genomics) about the feasibility of predicting ARGs from readily-available antimicrobial susceptibility profiles of bacteria generated by diagnostic laboratories. ARG prediction models based on random forests, support vector machines and generalized linear models were trained on an extensive collection of clinically relevant bacteria with diverse antibiotic susceptibility profiles. Model performance evaluation using leave-one-out cross validation suggests that support vector machines outperforms other methods for this task.

Rationale. Antibiotic resistance of nosocomial bacteria progresses at an alarming rate, menacing essential medical advances. To combat resistance with efficient antibiotic stewardship and infection control (AS/IC) interventions, we must gain a deeper understanding of the eco-evolutionary phenomena that govern the dissemination dynamics of resistant bacteria in hospital ecosystems. Integrating the phenomena most relevant to antibiotic resistance, including horizontal gene transfer and competition between bacterial lineages, would allow to model, understand and predict the outcome of interventions against resistance with unprecedented precision.

Objectives. We will build and inform a multiscale, stochastic model of the bacterial ecosystem in the 5,400-bed Hospices Civils de Lyon (HCL) teaching hospital group. Model parameters will be calibrated using massive microbiological and genomic data. After model calibration, we will simulate an exhaustive set of AS/IC interventions. Cost-effectiveness analysis of intervention outcomes will identify candidate optimal strategies against resistance and, importantly, will clarify the eco-evolutionary phenomena behind their optimality.

Methods. Our ecosystem model structure will consist of a hierarchy of nested elements, including antibiotic resistance-conferring genes, transposons, plasmids, bacterial cells, patients, and hospital wards. Stochastic events will capture the dynamics of transmission and movement of genes, bacteria and patients, as well as bacterial population dynamics and competition under varying antibiotic pressure regimes. Model parameters will derive, using Bayesian methods, from 5 years of exhaustive data on microbiological diagnoses, antibiotic prescriptions and patient trajectories, provided by the ongoing project EpiTrack conducted at the HCL. Information on the resistome and mobilome of ~100,000 bacterial isolates will be inferred from standard-of-care microbiological data by constructing a probabilistic resistome inference model, informed by whole-genome sequence analysis of a carefully selected subset of ~1,600 isolates representative of the diversity of resistance profiles in the HCL. A subset of parameters will be actionable to represent AS/IC interventions, such as targeted antibiotic restrictions or global hygiene improvements. Interventions will be simulated as variations of actionable parameters and their predicted outcomes will be monitored and compared.

Expected results and impact. ResisTrack will generate a holistic, bottom-up description of the dynamics of antibiotic resistance and of the influence of antibiotic prescription habits and patient movements. Simulation results will allow to propose integrated AS/IC strategies with maximal expected benefit against resistance, to be tested in field conditions in a follow-up multicentric study. The probabilistic resistome inference models will allow other clinical microbiology labs to monitor the dissemination of antibiotic resistance genes and mobile genetic elements in hospitals and the community at no added costs, using only standard-of-care microbiological data.

Consortium. The project will be conducted by French research centers with complementary skills in clinical microbiology, evolution, ecology and bioinformatics. Data integration, strain selection, DNA sequencing and model design will be performed at the Centre International de Recherche en Infectiologie of Lyon (CIRI Inserm U1111), in close collaboration with the Institute for Infectious Agents of the HCL. Bioinformatic analyses will be supervised by G. Perrière at LBBE/PRABI, Lyon (CNRS). T. Wirth at EPHE, Paris (Muséum National d’Histoire Naturelle), will provide expertise in bacterial eco-evolution and Bayesian population genetics methods. Mathematical modelling and project coordination will be led by J.P. Rasigade, principal investigator of the EpiTrack project already funded by the HCL and the FINOVI Foundation.