Inference of Infectious Disease Dynamics from Genetic Data via Sequential Monte Carlo

Abstract

When an epidemic moves through a population of hosts, the process of transmission may leave a signature in the genetic sequences of the pathogen. Patterns in pathogen sequences may therefore be a rich source of information on disease dynamics. Genetic sequences may replace or supplement other epidemiological observations. Furthermore, sequences may contain information not present in other datatypes, opening the possibility of inferences inaccessible by other means. The field of phylodynamic inference aims to reconstruct disease dynamics from pathogen genetic sequences.

Although a wide variety of phylodynamic inference methods have been proposed, most methods for fitting mechanistic models of disease operate in two disjoint steps, first estimating the phylogeny of the pathogen and then fitting models of disease dynamics to properties of the estimated phylogeny. Logical inconsistency in demographic assumptions underlying the two stages of inference may create bias in resulting parameter estimates. Joint inference of disease dynamics and phylogeny ensures consistent assumptions, but few methods for joint inference are currently available. The central work of this thesis is a new method for joint inference of disease dynamics and phylogeny from pathogen genetic sequences. This likelihood-based method, which we call genPomp, allows for fitting mechanistic models of arbitrary complexity to genetic sequences.

The organization of this thesis is as follows. In Chapter I, we present background on the field of phylodynamic inference. In Chapter II, we use simulation to study a two-stage inference approach proposed by Rasmussen et al. (2011). We find that errors in phylogenetic reconstruction may drive bias in two-stage phylodynamic inference. This result underscores the need for methodology for joint inference of the transmission model and the pathogen phylogeny. In Chapter III, we propose a flexible method for joint inference and demonstrate the feasibility of this method through simulation and a study on stage-specific infectiousness of HIV in Detroit, MI. This method is comprised of a class of algorithms that use sequential Monte Carlo to estimate and maximize likelihoods. In Appendix A we show theoretical support for our algorithms. In Chapter IV, we demonstrate the flexibility of our approach by developing a model of transmission of Vancomycin-resistant enterococcus in a hospital setting. To allow for fitting this model to patient-level data we developed a targeted proposal, detailed in Appendix B. We present exploratory analysis of a hospital outbreak at NIH that motivates the form of the model, and carry out a study on simulated data. Although some assumptions of the simulated example are unrealistic, these initial results will inform future efforts at fitting real data. In Chapter V, we summarize the progress represented in this thesis and consider possibilities for future work.