Improving Antimicrobial Resistance (AMR) Phenotype Prediction for Unseen Bacteria through Data Augmentation and Machine Learning

Abstract

Antimicrobial resistance (AMR) poses a growing threat to global health, yet accurately predicting resistance remains challenging due to limited and imbalanced genomic data. Existing machine learning (ML) models often struggle to generalize across diverse genome-antibiotic combinations, especially in data-scarce environments. AMR prediction is hindered by the high cost and time required for experimental data collection, as well as the lack of sufficient labeled genomic datasets. Additionally, current ML models face limitations in performance and generalizability when applied to small or imbalanced datasets. To address these challenges, we propose k-mer-based feature extraction for ML models. We further enhance model performance and generalizability through data augmentation techniques, enabling accurate AMR prediction despite limited training data.Our experimental results show that tree-based models with hyperparameter tuning perform well on small datasets, while CNN models trained with augmented data achieve superior accuracy and robustness across diverse genome-antibiotic pairs. Notably, the proposed deep 1D CNN framework achieves performance that equals or surpasses prior benchmarks, underscoring the benefit of coupling data augmentation with deep learning for AMR prediction. Furthermore, we validate that the proposed approach can successfully generalize to newly identified bacteria by accurately classifying multiple antibiotic classes in a multiclass prediction setting. These findings demonstrate that tree-based ensemble models enable rapid and reliable predictions on small datasets, while deep learning models achieve higher performance by leveraging data augmentation to address data scarcity. This study establishes a foundation for reducing antibiotic misuse by enabling accurate antibiotic selection, even for newly identified and previously unseen bacterial strains.