Because virus evolution and transmission often unfold on similar time scales, the shape of a virus phylogenetic tree can contain significant information about its epidemiology. Fitting epidemic models to tree shapes has generally focused on the distribution of node heights that are modelled through coalescent or birth-death processes. The combination of kernel methods and approximate Bayesian computation (kernel-ABC) provides a new, versatile approach to phylodynamic inference utilizing a richer source of information on tree shapes.
Parameter estimation by ABC is carried out by simulating data sets under different parameterizations of the model. Simulated trees are compared to the observed tree by a similarity measure that is provided by the kernel method. By rejecting parameter proposals on the basis of this similarity measure, the target distribution approximates the posterior. Thus, the range of epidemic models that can potentially fit with kernel-ABC is limited only by the ability to simulate trees.
First, I will review previous work using a Markov chain Monte Carlo (MCMC)-based implementation of ABC to fit conventional birth-death and coalescent models to tree shapes, and some of the limitations of ABC-MCMC in this context. Next, I will present current work using ABC with sequential Monte Carlo (SMC) to fit a contact network model to tree shapes. Networks are simulated from the Barabási-Albert (BA) preferential attachment model. Transmission trees are simulated from a network using a standard Gillespie algorithm. Parameters of the BA model are not only associated with variation in tree shapes, but are also identifiable from the shape of a given tree. I will discuss the implications of these results for our understanding of epidemics, and potential directions of research with kernel-ABC methods with respect to microbial source attribution.