Deep learning from phylogenies to uncover the transmission dynamics of epidemics

Voznica, J; Zhukova, A; Boskova, V; Saulnier, E; Lemoine, F; Moslonka-Lefebvre, M; Gascuel, O

Widely applicable, accurate and fast inference methods in phylodynamics are needed to fully profit from the richness of genetic data in uncovering the dynamics of epidemics. Standard methods, including maximum-likelihood and Bayesian approaches, which are both model specific, often rely on complex mathematical formulae and approximations, and do not scale well with dataset size. We develop a likelihood-free, simulation-based approach, which combines deep learning with (1) a large set of summary statistics measured on phylogenies or (2) a complete and compact vectorial representation of trees, which avoids potential limitations of summary statistics and applies to any phylodynamic model. Our method enables both model selection and estimation of epidemiological parameters. We demonstrate its speed and accuracy on simulated data, where it performs better than the state-of-the-art methods. To illustrate its applicability, we assess the dynamics induced by superspreading individuals in an HIV dataset of men- having-sex-with-men in Zurich.