Phylogenies as graphs: structured neural networks improve host origin predictions from paramyxovirus sequences
Accurately identifying the host of a virus from its genome sequence is a task with important applications in zoonotic disease surveillance and filling data gaps for metagenomic sampling. Machine learning approaches have seen broad application in making host predictions directly from viral genome sequences. However, most host prediction models do not incorporate information on viral phylogeny, which is strongly correlated with both genome composition and host. We apply a novel graph neural network (GNN) approach which explicitly represents viral phylogeny in model architecture to predict hosts of origin within the paramyxoviruses. We conduct rigorous benchmarking against non-structured neural networks and predictions made using phylogeny alone, showing that GNNs carry distinct advantages over other methods when making predictions where training data is sparse. Validation across different phylogenetic scales shows that simple phylogenetic prediction is effective in many applications and that phylogeny contributes a large proportion of the predictive power of host prediction models, with genome compositional features providing additional power only for specific predictions outside the range of the training data. This novel modelling approach and model validation framework are flexible and can be applied to other viral families.