Fig. 2

Host prediction performance of Phirbo, BLAST, and WIsH. The performance is provided by receiver operating characteristic (ROC) and precision–recall (PR) curves and statistical measures (i.e., F1 score, precision, and recall) separately for (a) Edwards et al. and (b) Galiez et al. data sets. ROC curves and the corresponding area under the curve (AUC) display the classification accuracy of virus–host predictions across all possible virus–prokaryote pairs. Dashed lines represent the levels of discrimination expected by chance. Dashed lines in the PR curve plots represent the levels of discrimination expected by chance. Score cut-offs for each tool were set to ensure the highest F1 score. (c), (d) Number of correctly predicted virus–host interactions (%) in the Edwards et al. and Galiez et al. data sets, respectively. Bars indicate the number of viruses for which a correct host was predicted at the species (blue bars) and genus (red bars) levels out of all phages in Edwards et al. (n = 820) and Galiez et al. (n = 1,420)