From sequence to signature: Machine learning uncovers multiscale feature landscapes that predict AMR across ESKAPE...

Area of Science:

• Genomics and Computational Biology
• Infectious Diseases and Microbiology
• Machine Learning in Healthcare

Background:

• Antimicrobial resistance (AMR) poses a significant global health challenge, with pathogens developing resistance faster than new drugs are discovered.
• Traditional methods for detecting AMR are slow and resource-intensive, necessitating advanced computational approaches for rapid prediction.
• Genomic sequencing generates vast amounts of data, creating opportunities for machine learning (ML) to predict AMR phenotypes and mechanisms.

Purpose of the Study:

• To develop and validate a comprehensive multiscale machine learning (ML) approach for predicting antimicrobial resistance (AMR) phenotypes.
• To identify molecular features (genes, proteins, domains) associated with drug-specific or drug-class-specific AMR.
• To provide a tool for reliable prediction of AMR in newly sequenced pathogen genomes and elucidate underlying resistance mechanisms.

Main Methods:

• Utilized sequenced genomes with experimentally derived AMR phenotypes for a subset of ESKAPE pathogens.
• Constructed pangenomes, clustered sequences, and extracted protein domains to generate multiscale features.
• Trained logistic regression ML models to predict AMR and identify associated molecular features, evaluating performance across scales, data types, drugs, and geographical/temporal holdouts.

Main Results:

• The ML models achieved high performance in predicting AMR phenotypes, with a median normalized Matthews correlation coefficient of 0.89.
• Identified known and novel AMR-associated genes, proteins, and domains, providing candidates for experimental validation.
• Demonstrated model resilience across geographical and temporal variations, and successfully uncovered multi-drug class resistance features.

Conclusions:

• The developed multiscale ML approach offers a reliable method for predicting existing and emerging AMR in bacterial pathogens.
• The approach effectively pinpoints the molecular contributors to AMR, aiding in understanding resistance mechanisms.
• The study provides an interactive web application for accessing models and results, facilitating further research and application in combating AMR.