Horizontal gene transfer (HGT) accelerates the spread of antimicrobial resistance (AMR) via mobile genetic elements allowing pathogens to acquire resistance genes across species. This process drives the evolution of multidrug-resistant “superbugs” in clinical settings. Detection of HGT is critical to mitigating AMR, but traditional methods based on sequence assembly or comparative genomics lack resolution for complex transfer events. While machine learning (ML) promises improved detection, several studies in other domains have demonstrated that data representations will strongly influence its performance. There is, however, no clear recommendation on the best data representation for HGT detection. Here, we evaluated 44 genomic data representations using five ML models across four data sets. We demonstrate that ML performance is highly dependent on the genomic data representation. The RCKmer-based representation (k = 7) paired with a support vector machine is found to be optimal (F1: 0.959; MCC: 0.908), outperforming other approaches. Moreover, models trained on multi-species data sets are shown to generalize better. Our findings suggest that genomic surveillance benefits from task-specific genome data representations. This work provides state-of-the-art, fine-tuned models for identifying and annotating genomic islands that will enable proper detection of transfer of AMR-related genes between species.
