Predicting S. aureus antimicrobial resistance with interpretable genomic space maps.

Area of Science:

• Genomics
• Machine Learning
• Computational Biology

Background:

• Antimicrobial resistance (AMR) is a critical global health challenge requiring faster methods for selecting effective antibiotic treatments.
• Genomic data combined with machine learning (ML) offers potential for rapid prediction of bacterial resistance phenotypes, but many ML models lack interpretability.

Purpose of the Study:

• To introduce a novel methodology for visualizing genomic sequence space and predicting AMR using Generative Topographic Mapping (GTM).
• To enhance the interpretability of ML models for AMR prediction.

Main Methods:

• Applied Generative Topographic Mapping (GTM), a non-linear dimensionality reduction technique, to analyze genomic data from over 5000 Staphylococcus aureus isolates.
• Developed GTM models for predicting antibiotic resistance phenotypes and visualizing the genomic space associated with AMR.

Main Results:

• GTM models achieved reasonable accuracy for predicting resistance across all tested antibiotics, with balanced accuracy values of at least 0.75.
• The generated Generative Topographic Maps (GTMs) effectively visualized genomic space, enabling antibiotic-wise comparison of resistant phenotypes.
• The GTM approach facilitated the analysis of genetic factors contributing to drug resistance.

Conclusions:

• The GTM-based methodology provides a valuable tool for exploring genomic sequence space and predicting AMR.
• This approach enhances the interpretability of ML models, aiding in the selection of optimal antibiotic treatments and understanding resistance mechanisms.