For the past 25 years, one discipline has been revolutionizing epidemiology: phylodynamics. By using genetic sequences, scientists can study how pathogens evolve. Thanks to mathematics, these data can be coherently organized into phylogenetic trees, with each branch representing a chain of transmission and each leaf representing an individual’s infection. The length of the branches can provide information about the number of mutations, while the branching points reveal the phylogenetic relationships between pathogens. Thus, a given strain can be traced back to its parent strain to compare their genetic sequences.
These phylogenetic trees are built using algorithms, which suggest the most likely tree based on the available genetic data. Various parameters can be estimated, such as the rate at which a pathogen spreads, the duration of infection, its geographical origin, and so on.
To produce robust estimates, a large number of genetic sequences must be analysed and the estimated parameters compared with conventional epidemiology data, which are based on the observation of populations of individuals.
This requires scaling up mathematical models, which are currently limited by data complexity (combination of genetic sequences with dates, prevalence rates 1 and geographical locations, etc.) and the volume of data to be processed.
Scientists at INRAE used a deep learning artificial intelligence method known as neural posterior estimation (NPE), which is typically applied in neuroscience and astrophysics. Using a dataset comprising 72 genomes from the 2014 Ebola outbreak in Sierra Leone, the team demonstrated the robustness of this method. The parameters estimated from the phylogenetic trees are similar to those obtained using traditional inference methods. This is the first time this method has been used with genetic data in a phylodynamic application.
NPE enables mathematical models to be calibrated much more quickly and to incorporate a large volume of various types of data (potentially thousands of sequences). The use of powerful algorithms paves the way for real-time management of epidemics and epizootics, with live scenario testing and the ability to add new data at any time. The scientists have made detailed online tutorials available to enable others to replicate this type of analysis .
Proceedings of the Royal Society B Biological Sciences
Simulation-based inference of epidemiological and phylodynamic models via neural posterior estimation
26-Aug-2026