Researchers at INRAE have demonstrated that a deep-learning technique known as neural posterior estimation (NPE) can significantly accelerate the mathematical analysis used to understand how infectious diseases spread and evolve.
The approach could eventually help scientists track epidemics and animal disease outbreaks in near real time, allowing them to test possible scenarios as new information becomes available.
The research builds on the growing field of phylodynamics, which for about 25 years has combined genetics, epidemiology and mathematics to reconstruct the history and spread of pathogens.
By analysing the genetic sequences of viruses and other pathogens, scientists can determine how they are related and reconstruct possible chains of transmission. The information is commonly represented through phylogenetic trees, mathematical structures in which branches illustrate relationships between pathogens and leaves correspond to individual infections or sampled cases.
The length and structure of the branches can provide clues about how pathogens have changed over time. Mutations accumulating in genetic sequences can help researchers determine how closely related different strains are, while branching points can indicate when one lineage diverged from another.
This makes it possible to trace a pathogen backwards through its evolutionary history and compare a particular strain with its likely parent strain.
However, building and analysing these trees becomes increasingly difficult as the amount and variety of available data grow.
Researchers must combine genetic sequences with information such as sampling dates, infection prevalence and geographical locations. They must then use mathematical models to estimate critical epidemiological parameters, including how rapidly a pathogen spreads, how long infections last and where an outbreak may have originated.
For these estimates to be reliable, scientists need to analyse large numbers of genetic sequences and compare the results with conventional epidemiological information collected from affected populations.
The sheer scale and complexity of such datasets, however, can make conventional mathematical approaches computationally demanding.
To address the challenge, scientists at INRAE applied neural posterior estimation, a deep-learning artificial intelligence technique more commonly used in fields such as neuroscience and astrophysics.
The researchers tested the method using genetic data from the 2014 Ebola outbreak in Sierra Leone. Their dataset contained 72 viral genomes, allowing the team to assess whether the AI-based approach could produce reliable estimates from phylogenetic information.
The results were encouraging.
Parameters estimated from the phylogenetic trees using NPE were similar to those obtained through more traditional statistical inference methods, demonstrating that the AI technique could reproduce reliable epidemiological estimates.
The study, published in the journal Proceedings B, represents the first reported application of the method to genetic data in a phylodynamic setting.
A major advantage of NPE is its ability to calibrate mathematical models much more rapidly than conventional approaches. It can also accommodate large and diverse datasets, potentially involving thousands of genetic sequences alongside epidemiological and geographical information.
The researchers say this could significantly expand the ability of scientists to monitor infectious diseases as they spread.
Instead of waiting until an outbreak has largely unfolded before conducting complex analyses, researchers could potentially update their models continuously as new genetic sequences and epidemiological information become available.
The technology could therefore support live scenario testing, allowing scientists and public health authorities to examine how an outbreak might develop under different circumstances and assess possible intervention strategies.
Beyond human diseases, the approach could also be applied to epizootics, or disease outbreaks among animal populations.
The researchers have also made detailed online tutorials available, providing guidance that could allow other scientists to reproduce and build on the analysis.
As genetic sequencing becomes increasingly accessible and disease surveillance produces ever-larger datasets, AI-assisted phylodynamics could become an important tool for understanding outbreaks faster and improving the speed of epidemic response.