Estimation of country-specific tuberculosis antibiograms using genomic data

Avika Dixit, Luca Freschi, Roger Vargas, Matthias Groeschel, Sabira Tahseen, Sk Mahbub Alam, S. M. Mostofa Kamal, Alena Skrahina, Ramon Basilio, Dodge R. Lim, Nazir Ismail, Maha FarhatmedRxiv 2021 · DOI

Abstract

Global tuberculosis (TB) drug resistance (DR) surveillance is largely focused on rifampicin, leaving significant gaps in our understanding of resistance to other critical antibiotics. This study leverages public and surveillance M. tuberculosis (Mtb) whole genome sequencing (WGS) data to generate comprehensive country-level resistance prevalence estimates (antibiograms) for 12 drugs. By applying a validated random forest model and correcting for sampling biases—including outbreak clusters and resistance oversampling—we estimated resistance profiles for 29 countries. Our findings reveal high rates of fluoroquinolone resistance in South Asia and suggest that ethionamide remains a significantly underutilized drug for treating isoniazid-resistant TB.

Introduction

Current global TB surveillance is "rifampicin-centric." Because traditional culture-based drug susceptibility testing (DST) is slow and resource-intensive, international monitoring often relies on rifampicin resistance as a proxy for multi-drug resistance (MDR). This approach creates a "black box" regarding the effectiveness of other first- and second-line drugs. While Whole Genome Sequencing (WGS) offers a rapid alternative for resistance prediction, public genomic databases are inherently biased; researchers typically sequence isolates from major outbreaks or highly resistant cases rather than representative national samples. This study introduces a framework to de-bias public genomic data, providing the first global attempt at estimating comprehensive Mtb antibiograms.

Methods

To transform biased genomic data into reliable prevalence estimates, we developed a multi-step analytical pipeline:

  1. Genomic Prediction: We utilized a validated Random Forest (RF) classifier to predict phenotypic resistance based on mutations across 18 genetic loci.
  2. Outbreak Correction: To prevent localized outbreaks from skewing national data, we calculated pairwise Single Nucleotide Polymorphism (SNP) distances and pruned closely related isolates (distance ≤ 10 SNPs).
  3. Bayesian Bias Correction: We applied a Bayesian model to adjust for the imperfect sensitivity and specificity of the RF model, ensuring the estimates reflected true phenotypic resistance.
  4. Resistance Re-weighting: We used official WHO rifampicin resistance rates as a "ground truth" anchor. By re-weighting the oversampled MDR isolates in the genomic datasets, we normalized the data to reflect the actual national burden of each country.

Our primary dataset included 22,599 Mtb genomes curated from NCBI, PATRIC, and published literature. We validated these estimates using national drug resistance survey data from South Africa (2012–2014).

Key Results and Validation

The framework was applied to 29 countries that met the quality threshold of 100+ sequenced isolates. Validation against the South African national survey showed that marginal genotypic resistance estimates overlapped with survey results for nearly all drugs, though isoniazid and second-line injectables were slightly underestimated. Among MDR-TB cases, estimates for pyrazinamide and fluoroquinolones were highly accurate.

Globally, isoniazid mono-resistance was estimated at 10.9% (95% CI: 10.2-11.7%). Regional hotspots were prominent in former Soviet Union countries; for example, Russia was estimated to have a 66% isoniazid resistance rate.

Clinical and Policy Implications

The South Asia Fluoroquinolone Warning

One of the most critical findings concerns fluoroquinolone (FQ) resistance. In India and Pakistan, mono-levofloxacin resistance rates were estimated at 2.8% and 3.4%, respectively, among rifampicin-susceptible cases. This is particularly concerning as new, shorter 4-month treatment regimens rely heavily on FQs. If resistance is already present in supposedly "drug-susceptible" cases, these new regimens may face premature failure in high-burden regions.

Ethionamide Underutilization

The study revealed that 74.4% (IQR: 64.5-79.7%) of isoniazid-resistant isolates were predicted to be susceptible to ethionamide. Ethionamide is frequently sidelined in clinical practice due to its side-effect profile, but these results suggest it is a powerful, underused tool for treating isoniazid-resistant or MDR-TB.

Reality Check for Short-Course Regimens

The "Bangladesh regimen" for MDR-TB requires susceptibility to both pyrazinamide and levofloxacin. However, our global estimate found that only 15.1% (95% CI: 10.2-19.9%) of MDR-TB isolates were susceptible to both drugs. This suggests that the utility of such empiric short-course regimens may be more limited than previously hoped, necessitating more personalized treatment approaches.

Limitations

Despite the progress, geographic coverage remains a challenge; many high-burden countries are "genomic deserts" with insufficient data for reliable estimation. Furthermore, the current model cannot predict resistance to newer cornerstone MDR-TB drugs like bedaquiline or linezolid due to a lack of sufficient training data. Finally, the lack of clinical metadata in public databases (e.g., whether a patient is new or previously treated) remains a barrier to even more granular modeling.

Conclusion

This research demonstrates that public WGS data, when properly de-biased, can serve as a reliable proxy for expensive field surveys. By providing a more nuanced view of the TB resistance landscape, these genomic antibiograms can help clinicians and policymakers optimize treatment regimens and improve global TB control efforts. The results are accessible via an interactive web platform at GenTB.

Summary updated

Abstract

Abstract Background Global tuberculosis (TB) drug resistance (DR) surveillance is largely focused on the drug rifampicin. We leveraged public and surveillance M. tuberculosis ( Mtb) whole genome sequencing (WGS) data, to generate more comprehensive country-level resistance prevalence estimates (antibiograms) using in silico resistance prediction. Methods We curated and quality-controlled Mtb WGS data. We used a validated random forest model to predict phenotypic resistance to twelve drugs and bias-corrected for model performance, outbreak sampling, and resistance oversampling. We validated our estimates using a national DR survey conducted in South Africa. Results Mtb isolates from 29 countries (n=19,149) met sequence quality criteria. Marginal genotypic resistance estimates overlapped with the South African DR survey for all drugs except isoniazid and second-line injectables that were underestimated (n=3,134); among multi-drug resistant (MDR) TB, estimates overlapped for pyrazinamide and the fluoroquinolones. Globally, mono-resistance to isoniazid was estimated at 10.9% (95% CI: 10.2-11.7%, n = 14,012. Mono-levofloxacin resistance rates were highest in South Asia (Pakistan 3.4% [0.1-11%], n=111 and India 2.8% [0.08-9.4%], n=114). Rates of resistance discordance between isoniazid and ethionamide were high with 74.4% (IQR: 64.5-79.7%) of isoniazid resistant isolates predicted to be ethionamide susceptible. The global susceptibility rate to pyrazinamide and levofloxacin among MDR was 15.1% (95% CI: 10.2-19.9%, n=3,964). Conclusions This is the first attempt at global Mtb antibiogram estimation. DR prevalence in Mtb can be reliably estimated using public WGS and phenotypic resistance prediction for key antibiotics. Our results raise concerns about the empiric use of short-course fluoroquinolone regimens for drug susceptible TB in South Asia and suggest that ethionamide is an under-utilized drug in MDR treatment.
BibTeX
@misc{dixit2021estimation,
  author = {Avika Dixit and Luca Freschi and Roger Vargas and Matthias Groeschel and Sabira Tahseen and Sk Mahbub Alam and S. M. Mostofa Kamal and Alena Skrahina and Ramon Basilio and Dodge R. Lim and Nazir Ismail and Maha Farhat},
  title = {{Estimation of country-specific tuberculosis antibiograms using genomic data}},
  year = {2021},
  doi = {10.1101/2021.09.23.21263991},
  howpublished = {medRxiv},
}
Figure