A statistical learning approach to infer transmissions of infectious diseases from deep sequencing data

National audiencePathogen sequence data have been exploited to infer who infected whom, by using empirical and model-based approaches. Most of these approaches exploit one pathogen sequence per infected host unit (e.g., individual, household, field). However, data collected with deep sequencing techniques, providing a subsample of the pathogen variants at each sampling time, are expected to give more insight on epidemiological links than a single sequence per host unit. A mechanistic viewpoint to transmission and micro-evolution has generally been followed to infer epidemiological links from these data. Here, we investigate an alternative statistical learning approach for estimating epidemiological links, which consists of learning the structure of epidemiological links with a pseudo-evolutionary model and training data before inferring links for the whole data set. We designed the pseudo-evolutionary model as a semi-parametric regression function where the response variable is the set of sequences observed from a recipient host unit and the explanatory variable is the set of sequences observed from a putative source. We derived from this model a penalized pseudo-likelihood that is used for selecting who infected whom or who is closely related to whom, where the penalization is calibrated on training data. In order to assess the efficiency of the pseudo-evolutionary model and the associated inference approach for estimating epidemiological links, we applied it to simulated data generated with diverse sampling efforts, sequencing techniques (corresponding to diverse depths and read lengths), and stochastic models of viral evolution and transmission. Then, we applied it to three real epidemics: swine Influenza, Ebola and a potyvirus of wild salsify. Such an approach has the potential to be particularly valuable in the case of a risk of erroneous mechanistic assumptions and sequencing errors, it is sufficiently parsimonious to allow handling big data sets in the future, and it can be applied to very different contexts from animal, human and plant epidemiology