UCLA

Phylodynamic inference exploits pathogen genomic data to estimate epidemiological quantities of direct public-health relevance, including time-varying transmission and recovery rates, effective population size trajectories, and migration rates between regions. Bayesian implementations on coalescent and birth–death tree priors are standard in outbreak surveillance, but the underlying inference machinery scales poorly with the size of contemporary genomic datasets and with the parameter dimensionality of modern episodic and structured models. This dissertation develops scalable algorithms, both gradient-based and parallel, that remove these bottlenecks for two model families central to applied phylodynamics. The first contribution targets the episodic birth–death–sampling model. The likelihood is reformulated to admit a closed-form, linear-time analytic gradient with respect to all epoch-specific birth, death, and sampling parameters, enabling Hamiltonian Monte Carlo sampling of the full epoch-rate vector. Combined with regularized shrinkage priors over the high-dimensional rate vectors, the resulting sampler increases the minimum effective sample size per unit time 10- to 200-fold over univariate Metropolis–Hastings, while recovering smooth effective-reproductive-number trajectories. The methodology is demonstrated on HIV-1 subtype A sequences from Odesa, Ukraine, a New York State seasonal influenza A/H3N2 hemagglutinin alignment, and the 2014–2016 West African Ebola virus epidemic. The second contribution addresses the structured coalescent approximation that underlies modern phylogeographic inference. Reformulating its peeling recursion to expose fine-grained parallelism across both lineages and demes accelerates the likelihood by 10–26-fold on multi-core central processing units and up to 60-fold on graphics processing units, placing previously intractable analyses within practical wall-clock budgets. Reverse-mode differentiation of the same recursion then returns the exact gradient for all migration and population-size parameters at the cost of a few likelihood evaluations rather than one per parameter, making Hamiltonian Monte Carlo practical for high-dimensional phylogeographic posteriors. The methods are validated on a dengue virus phylogeographic analysis and applied to a continental-scale reconstruction of highly pathogenic avian influenza A(H5N1). A concluding chapter sketches extensions: a log-linear parameterization of migration rates for hypothesis-driven analysis of pathogen spread, and both parametric and non-parametric models for time-varying effective population sizes. All methods are released as open-source extensions to the BEAST X phylogenetic inference platform and the BEAGLE library.