Assessing computational predictions of antimicrobial resistance phenotypes from microbial genomes

Area of Science:

• Genomics
• Computational Biology
• Microbiology

Background:

• Rapid whole-genome sequencing enables computational prediction of antimicrobial resistance (AMR) phenotypes.
• Both rule-based and machine learning (ML) methods exist, but systematic benchmarking is lacking.

Purpose of the Study:

• To rigorously evaluate and benchmark state-of-the-art ML methods against a rule-based method for AMR phenotype prediction.
• To identify the most effective computational approaches across diverse bacterial species and antibiotic combinations.

Main Methods:

• Evaluated four ML methods (Kover, PhenotypeSeeker, Seq2Geno2Pheno, Aytan-Aktug) and the rule-based ResFinder.
• Trained and tested methods on 78 species-antibiotic datasets using integrated evaluation approaches and sample splitting methods.

Main Results:

• Significant performance variation observed across methods and datasets; ML methods excelled on closely related strains, ResFinder on divergent genomes.
• Kover, PhenotypeSeeker, and Seq2Geno2Pheno were top-performing ML methods. Macrolides and sulfonamides showed highest prediction accuracies.
• Prediction accuracy varied by species-antibiotic combination, with beta-lactams and tetracyclines showing more variability. Campylobacter jejuni and Enterococcus faecium phenotypes were predicted most robustly.

Conclusions:

• No single method universally outperforms others; optimal choice depends on the specific species-antibiotic combination and genomic divergence.
• Highlights the need for method optimization for clinical applications, especially for divergent strains.
• Provides software recommendations for specific species-antibiotic pairings to guide practical implementation.