Model performance across all scheme scenarios.

Column definitions: AUC score; balanced accuracy score; Precision; Recall; species: name of the species that the model trained on; antibiotic: name of the antibiotic that the model trained on; MOA: mechanism of action for the antibiotic; case: testing scenario, corresponding to Fig 1. For example, A-a means scheme A training scenario a; R1: in scheme A, this denotes the number of resistant genomes in clade 1. In scheme B, it denotes the number of resistant genomes in the other clades where resistant or susceptible genomes were not excluded; S1: in scheme A, this denotes the number of susceptible genomes in clade 1. In scheme B, it denotes the number of susceptible genomes in the other clades where resistant or susceptible genomes were not excluded; Rn: in scheme A, this refers to the number of resistant genomes in the paired clades. In scheme B, it refers to the number of resistant genomes in the clade from which resistant or susceptible genomes were excluded; Sn: in scheme A, this refers to the number of susceptible genomes in the paired clades. In scheme B, it refers to the number of susceptible genomes in the clade from which resistant or susceptible genomes were excluded; distance: clade-wise distance (See “Applying machine learning to interpret model performance” in Methods); model: training model type; scheme: training scheme; split: sampling methods used for calculating the scores are indicated. ‘Scheme’ refers to sampling according to the defined scheme scenarios, while ‘Random’ refers to random sampling. See Training samples in Methods for details; drug_class: drug classes according to CARD. (XLSX)