Real-time reconstruction of epidemic dynamics from viral phylogenies using Deep Learning

ongoing

Reconstructing epidemic dynamics in real-time has become crucial for effective disease management, as demonstrated by the COVID-19 pandemic. Traditional methods rely on epidemiological data (e.g., reported cases), which can be biased or incomplete due to variable testing policies, particularly in resource-limited settings. Instead, phylodynamics has emerged as a valuable toolkit for using viral phylogenies to understand epidemic dynamics. However, conventional phylodynamic methods rely on mathematical formulas and approximations, which are not scalable to large datasets and are time-consuming, limiting their use primarily to retrospective rather than real-time analysis. This proposal aims to transform phylodynamics by integrating it with deep learning to bypass the cumbersome likelihood calculations, thereby facilitating real-time analysis directly from sequence data. I will develop innovative deep learning models to explore the relationships between phylogenetic trees and epidemiological parameters of viruses with epidemic potential, such as SARS-CoV-2, influenza, RSV, and mpox. These models are designed to rapidly and accurately estimate time-varying epidemiological metrics, including transmission heterogeneity, basic reproduction numbers, infectious and incubation periods. This initiative is set to revolutionize our ability to model and comprehend infectious diseases in real-time, elevating sequence data to a critical, standalone data source. It will incorporate cross-validation with epidemiological inference from reported cases and wastewater analyses, reducing reliance on any single data source and enhancing both public health responses and infectious disease surveillance. Furthermore, by simulating incremental data collection that reflects real outbreak conditions, this project will evaluate the sensitivity of real-time estimations and determine the necessary sampling proportions to accurately represent epidemic dynamics. This approach will yield crucial