Predicting antimicrobial resistance using conserved genes.

Area of Science:

• Genomics
• Machine Learning
• Microbiology

Background:

• Machine learning models accurately predict antimicrobial resistance (AMR) from bacterial sequence data.
• Current models often require complete genomes or extensive gene sets, limiting their use with incomplete data.

Purpose of the Study:

• To investigate the feasibility of predicting AMR phenotypes using incomplete bacterial genome sequences.
• To develop and evaluate machine learning models based on limited sets of conserved non-AMR genes.

Main Methods:

• Models were constructed using small, randomly selected sets of conserved core genes, excluding AMR genes.
• Predictive performance was assessed for *Klebsiella pneumoniae*, *Mycobacterium tuberculosis*, *Salmonella enterica*, and *Staphylococcus aureus*.
• Stability and variance of model performance were evaluated through random gene set sampling.

Main Results:

• Accurate prediction of susceptible and resistant phenotypes was achieved with as few as 100 conserved non-AMR genes (average F1 scores: 0.80-0.89).
• Very major error rates ranged from 0.11-0.23, and major error rates ranged from 0.10-0.20.
• Models demonstrated predictive power for AMR mechanisms involving single nucleotide polymorphisms (SNPs) and horizontal gene transfer, with stable performance metrics.

Conclusions:

• Sufficient genetic variation exists within core non-AMR genes for predicting AMR phenotypes, even with incomplete genome data.
• Machine learning models trained on limited core gene sets offer a promising approach for AMR prediction when complete genomes are unavailable.
• This method provides a robust alternative for AMR surveillance and diagnostics in resource-limited settings.