Teddy: neural inference of epidemiological parameters from viral sequences

Abstract

Estimating how fast infections spread or how long they last is essential to control outbreaks. Phylodynamics methods enable the inference of key epidemiological pa-rameters from viral genomic data but remain limited in terms of biological realism and speed because they need to derive and compute likelihoods. We address this is-sue using simulation-based inference and introduce a deep learning-based framework directly trained on alignments of viral genetic sequences. Our neural posterior estima-tion matches the accuracy of leading Bayesian likelihood-based methods while running a thousand times faster and avoiding a phylogeny reconstruction step. This perfor-mance can be harnessed to analyze large datasets and opens new perspectives to tackle biologically realistic models in terms of pathogen life histories or genomic evolution.