Machine learning methods in predicting antimicrobial resistance

Abstract

Antimicrobial resistance poses a significant global threat to healthcare systems worldwide. The high cost and short lifespan of new antibiotics due to the rapid evolution of resistance exacerbate this crisis, prompting exploration of diverse strategies to combat AMR. Within the context of AMR diagnostics, machine learning approaches for predicting bacterial resistance are under active investigation. This paper reviews state-of-the-art methods for predicting antibiotic resistance in individual bacterial isolates. The review encompasses both traditional ML algorithms, such as logistic regression and random forests, and more advanced techniques including convolutional and graph neural networks, as well as transformer language models. The inherent high variability of bacterial genomes and the heterogeneity of mutational processes hinder model generalizability, while variations in clinical conditions during strain development contribute to prediction instability on new datasets. This necessitates further research into improved data collection and preprocessing methods, as well as a more comprehensive consideration of both known and emerging resistance mechanisms.