Use the following steps to reproduce any dataset that was used in the paper.
The datasets are organized by species and antibiotic. Each directory contains a metadata file with two columns: 1) the genome identifiers (PATRIC) and 2) the phenotypes.
To download all the genomes in a dataset, simply feed the corresponding metadata.tsv file to the download_genomes.py script. For example:
python download_genomes 'mycobacterium tuberculosis/kanamycin/metadata.tsv' genomes_dir
Now genomes_dir contains the genomes and a summary file called genome_paths.tsv, containing the path to each genome, has been created in the current directory. You will need this file in the next step.
Follow this tutorial, but replace the label file by the metadata.tsv that you just used and the genome file by the corresponding genome_paths.tsv.
To obtain the exact same datasets as in the paper, create 10 splits with 80% of the data for training, 10 folds, and random seeds 0, 1, 2, 3, 4, 5, 6, 7, 8, 9.
These datasets were extracted from the PATRIC database (https://www.patricbrc.org/). See our paper for the data acquisition protocol. More info on this database can be found in:
Antonopoulos, D. A., Assaf, R., Aziz, R. K., Brettin, T., Bun, C., Conrad, N., ... & Kenyon, R. W. (2017). PATRIC as a unique resource for studying antimicrobial resistance. Briefings in bioinformatics.
Wattam, A. R., Abraham, D., Dalay, O., Disz, T. L., Driscoll, T., Gabbard, J. L., ... & Machi, D. (2013). PATRIC, the bacterial bioinformatics database and analysis resource. Nucleic acids research, 42(D1), D581-D591.