Scalable and Epidemiologically Interpretable Phylodynamics to Recover Heterogeneous Transmission Dynamics

Project Summary As pathogens are transmitted between individuals, they accumulate mutations, leaving a footprint of the transmission history in the pathogen genomes. Using phylogenetic methods, we can reconstruct the transmission history connecting individual cases from these genomes, by reconstructing the relationships of the pathogens. We can then infer population-level transmission dynamics, from the ancestral relationships of the pathogens, or phylogenies, using phylodynamic methods. Infectious disease transmission and disease burden are highly heterogeneous, differing between neighborhoods, across age, and socioeconomic groups, and racial and ethnic lines. This heterogeneity means that it is crucial to a) be able to illuminate differential disease burdens and b) account for these heterogeneities when modeling or forecasting infectious disease outbreaks. Traditional approaches based on reported caseloads are often insufficient for capturing the full scope of highly heterogeneous transmission dynamics. Phylodynamics offers a potential solution, as it infers transmission dynamics from the connectivity of cases, providing an opportunity to disentangle these complex patterns. However, limitations in our available toolbox prevent us from fully utilizing the vast availability of pathogen genomes to study these complex transmission dynamics, as current phylodynamic approaches suffer from multiple challenges. With the advent of widely available sequencing, phylodynamic tools are not computationally efficient enough to analyze the amounts of data generated at the granular scales crucial to understanding transmission dynamics. Additionally, the model parameters need to be epidemiologically interpretable to be actionable. In this project, we seek to address these two points by developing novel approaches to reconstruct transmission dynamics from pathogen sequence data. We will develop novel phylodynamic tools to reconstruct transmission dynamics at a granular scale by integrating neural networks into phylodynamic likelihood calculations that we show in preliminary results to dramatically improve computational efficiency and scalability. Phylodynamic methods are parameterized by more or less abstract parameters that either have no direct epidemiological meaning or are contingent on idealized assumptions about disease spread. We will establish how and when current approaches return biased results when reconstructing city-scale transmission dynamics, describe how they can be used to estimate actual disease burden, and test them using SARS-CoV-2 sequence data collected by Kaiser Permanente Southern California (KPSC) and in the UK over the pandemic. Finally, we will develop ways to quantify the factors influencing disease burden, such as geography, age, and socioeconomics. We will apply these tools to KPSC SARS-CoV-2 data, where we can access rich patient metadata to study these patterns. Our overarching goal is to utilize phylodynamic inference of heterogeneous transmission dynamics to parameterize complex infectious disease dynamic models and improve prediction accuracy.