Table 2.
Classification performance, all hosts. Whole available genomes. Negative class is the majority class. BAcc. – balanced accuracy, Rec. – recall, Spec. – specificity. BLAST (reads) and our networks use read-wise majority vote or output averaging to aggregate predictions over all reads from a genome. k-NN (genome) and BLAST (genome) use contig-wise majority vote. k-NN (contigs) and BLAST (contigs) represent performance on individual contigs treated as separate entities. k-NN (reads) was not used, as high conflicting prediction rates made read-wise aggregation impracticable
| Bacc. | AUPR | Rec. | Spec. | |
|---|---|---|---|---|
| CNNAll (ours) | 91.7 | 91.2 | 89.3 | 94.2 |
| LSTMAll (ours) | 86.3 | 85.8 | 96.2 | 76.4 |
| BLAST (reads) | 90.3 | n/a | 85.5 | 95.1 |
| k-NN (genome) | 82.8 | n/a | 93.9 | 71.6 |
| BLAST (genome) | 90.5 | n/a | 86.3 | 94.6 |
| k-NN (contigs) | 83.0 | n/a | 94.3 | 71.6 |
| BLAST (contigs) | 88.4 | n/a | 87.1 | 89.7 |