Erythromycin Resistance in S. aureus: DNABERT-2 vs XGBoost

Repository: nesirli/sa-amr-ml · Status: in progress. Results will be added here when the comparison is complete.

The question

Do DNA language models earn their compute cost for antimicrobial resistance prediction, or does a gradient-boosted model on known resistance genes do just as well?

Approach

• Data: S. aureus genomes from NCBI Datasets with erythromycin susceptibility phenotypes.
• Baseline: XGBoost on gene presence/absence features from AMRFinderPlus.
• Deep learning: DNABERT-2 fine-tuned with LoRA on sequence input.
• Comparison: the same split and the same metrics for both models, plus training and inference cost.