1653 - Estimation of country-specific tuberculosis antibiograms using a wide and deep neural net on a large genomic dataset

Background: Improved estimates of drug resistant tuberculosis (TB) burden are needed to aid control efforts. The World Health Organization (WHO) currently reports estimates for rifampin resistance (RR) or multidrug resistance (MDR) at the national level. Resistance rates to other first-line and second-line agents, e.g. ethambutol, pyrazinamide, and aminoglycosides, are rarely available, even at the country level. Our objective was to generate country and drug specific resistance prevalence estimates (antibiograms) using in silico phenotype prediction and curated public and surveillance Mycobacterium tuberculosis (MTB) genomic data.

Methods: We curated MTB genomes either by sequencing or from published literature and excluded genomes that did not meet our quality criteria (i.e. at least 10X depth in >95% of the genome). A machine learning model previously trained to predict phenotypic resistance in MTB with high accuracy, a wide and deep neural net (WDNN), was used to predict resistance to ten drugs. We corrected for resistance oversampling in genomic data by conditioning on RR and using country specific surveillance MDR/RR rates reported by the WHO.

Results: Of the 49,851 MTB genomes curated, 33,873 isolates met quality criteria. Of these, geographic data was available for 22,838 genomes. Antibiograms were generated for nine first- and second-line drugs for 36 countries. Among countries with at least 100 isolates, a high rate of resistance to fluoroquinolones and second line injectables was seen among isolates from the Republic of Moldova (15.4% [CI = 13.7-16.7%] moxifloxacin resistant, 6.3% [CI = 5.5-6.8%] kanamycin resistant, n = 330)  and Russian Federation (9.3% [CI = 9.1-9.4] moxifloxacin resistant, 5.4% [CI = 5.3-5.5%] kanamycin resistant, n = 1011) (Figure 1).Figure 1: Antibiograms created using genotypic data for isolates from Republic of Moldova (n=330, rifampin-resistance rate correction: 29%, range 26-31% among new tuberculosis cases);and Russian Federation (n=1011, rifampin-resistance rate correction 35%, range 34-35%, among new tuberculosis cases. rif: rifampin, inh: isoniazid, pza: pyrazinamide, emb: ethambutol, str: streptomycin, cap: capreomycin, amk: amikacin, kan: kanamycin, moxi: moxifloxacin


Conclusion: The estimation of antibiotic resistance prevalence in MTB for pyrazinamide, ethambutol and second-line agents can be aided by the use of in silico models of drug resistance. A high rate of resistance to second-line drugs precludes large scale roll out of short-course WHO regimens for treatment of MDR-TB for empiric use in certain countries. The use of whole genome sequencing for resistance surveillance can inform policy on optimal national regimen choice for TB treatment.