Bayesian latent class analysis of mobile genetic elements for identification of zoonotic spillover and host switch events

Abstract

Background: Extraintestinal pathogenic Escherichia coli causes millions of urinary tract infections and nearly 40,000 sepsis deaths each year in the United States. Accessory genetic elements may provide information regarding host origin of E. coli isolates. Mobile genetic elements, or clusters of genes, are labile and are preserved at differential rates between hosts. Detection of mobile genetic elements in conjunction with core-genome phylogenetics may identify zoonotic spillover and host switch events. A hierarchical Bayesian latent class model (HBLCM) is proposed to systematically integrate multiple accessory elements for probabilistic assignment of host-origins. Methods: The HBLCM assumes the host-origin for each isolate is in an unobserved class of human, turkey, pork and chicken, with further stratification into three major clades based on core phylogeny. HBLCM uses multivariate binary responses that indicate presence or absence of 18 host-associated accessory elements, identified previously, to infer the latent host-origins. The latent classes and model parameters can be learned in an unsupervised fashion or using a training set. Markov chain Monte Carlo algorithms are used to iteratively produce samples from the posterior distribution of the unobserved host-origins, based on which we calculate posterior probabilities of host-origins for each isolate. E.coli isolates (n=3,128) derived from human clinical isolates and turkey, chicken, and pork meat samples was randomly split 75/25 into training and validation sets.Results: HBLCM results matched human vs. meat host origins in 94.0% of isolates, and 74.7% of isolates from specific meat types in the validation set (n=782). Eight potential spillover events were identified a priori from phylogenetic trees built using the core genome. HBLCM results confirmed four as potential host switches, with three others demonstrating uncertainty. Conclusions: Mobile genetic elements and HBLCM constitute a principled approach for estimating probabilities of zoonotic spillover and host switch events and may inform policy and treatment recommendations.