The emergence of rapidly spreading viral variants (such as VOCs for COVID-19) can be decisive in the progression of epidemics, leading to new waves of infection and significant changes in viral populations. Phylodynamic methods can help identify and study the initial transmission dynamics of these variants. These methods exploit the fact that changes in virus population dynamics leave a phylogenetic signal to estimate evolutionary parameters. However, these methods assume that lineages are independent conditioning and, therefore, can only use local information to estimate the lineage’s dispersal potential. In this project, we postulate that the emergence of a new, rapidly dispersing lineage leaves a signature in the evolutionary history of all lineages, infecting the same population.

Furthermore, we postulate that this effect can be used to predict which lineages will become dominant. To this end, we propose adapting a new phylodynamic machine learning approach to the problem of identifying and tracking these variants. The machine learning approach will allow us to take advantage of the large number of molecular sequences available for these viruses, the analysis of which would be prohibitive for traditional phylodynamic methods. Furthermore, by better using global information, these methods may enable early detection of these variants, potentially even before substantial sampling of the emerging variant.

In this project, our objective is to study the influenza virus with the goal of making predictions that can guide strategies for managing the epidemic. We plan to characterize influenza from the perspective of molecular evolution and its interaction with the human immune system, also known as antigenic evolution. We aim to establish connections between these two processes. Viewing antigenic evolution through a clustering lens could potentially lead to improved predictions about the dominant strains in the upcoming year. The project has a significant methodological component, which involves developing suitable statistical techniques to handle high-dimensional data sets with complex dependency structures. Methodologically, we focus on two main points: firstly, we aim to develop a methodology based on U-statistics for inference in clustering problems. Secondly, we plan to develop Bayesian phylodynamic methods that explicitly model the relationship between molecular and phenotypic evolution, such as antigenic evolution.