Colloquium Speaker - Dr. Devan Becker, Wilfrid Laurier University

Abstract:

Wastewater-based epidemiology (WBE) for COVID-19 variants of concern ("variants") uses genetic sequencing of wastewater to find frequencies of individual "mutations", which are compared to known variants to determine their abundance. For WBE, variants are defined as a list of mutations. In a single sample this results in a constrained linear model, with covariates indicating the inclusion of a mutation in a variant's definition (1 if included, 0 otherwise; each mutation can belong to multiple variants) and coefficients corresponding to the abundance of that variant.

My research has focused on leveraging information from samples taken over space and time. If, say, Omicron comprises 40% of the wastewater sample on day t, then the abundance of Omicron on day t+1 should be similar. I tried using Time-Varying Coefficient models, but was very disappointed by the results (and someone else beat me to it, anyway). Instead of trying to make the variant abundances smooth in time, I model the individual mutations over time and then derive the variant abundances.

An additional issue comes from the definitions of variants themselves. They are derived from clinical sampling according to their placement on a phylogenetic tree, but wastewater deconvolution requires lists of mutations per variant. This mismatch causes many headaches. Due to the volume of sampling, I can instead find something that looks a lot like Omicron without a definition derived from clinical sampling, which ends up being a new type of clustering problem. Thanks to the magic of Bayesian estimation, I can pre-specify some known definitions and estimate the other definitions in a Bayesian Semi-Supervised Latent Overlapping Clustering Time-Varying Coefficients model, which is a fancy way to say that I am severely overthinking a basic linear model.