Estimation of country-specific tuberculosis resistance antibiograms using pathogen genomics and machine learning

Area of Science:

• Genomics
• Microbiology
• Epidemiology

Background:

• Global tuberculosis (TB) drug resistance (DR) surveillance primarily focuses on rifampicin resistance.
• Limited data exists on country-level prevalence of resistance to multiple TB drugs.

Purpose of the Study:

• To assess the utility of public whole-genome sequencing (WGS) data for predicting Mycobacterium tuberculosis (Mtb) drug resistance.
• To generate expanded country-level antibiograms using in silico resistance prediction.

Main Methods:

• Curated and quality-controlled Mtb WGS data from 29 countries.
• Employed a random forest model to predict phenotypic resistance to 12 drugs.
• Applied bias correction for model performance, sampling, and oversampling of rifampicin resistance.

Main Results:

• Public WGS data from 19,149 Mtb isolates were analyzed.
• In silico predictions showed overlap with South African DR survey data, with some drug under- or overestimations.
• High rates of isoniazid resistance (10.9%) and discordance between isoniazid resistance and ethionamide susceptibility (74.4%) were observed.

Conclusions:

• This study presents the first global Mtb antibiogram estimation using public WGS data.
• Public WGS data can reliably estimate DR prevalence but may oversample multidrug-resistant isolates.
• Findings raise concerns regarding fluoroquinolone use in South Asia and ethionamide underutilization in MDR-TB.