Graduate Researcher Tulane University New Orleans, Louisiana, United States
Introduction: Microbial drug resistance is a growing problem that can lead to prolonged treatment, increased healthcare costs, and risk for further drug resistance when treating microbial infections. Mycobacterium tuberculosis (Mtb) represents a striking example since Mtb infections can lead to tuberculosis, a leading cause of death from infectious disease that requires multi-drug treatment regimens that can lead to the emergence of multi-drug resistant and extremely drug-resistant strains. Although genome-wide association studies (GWAS) can identify genetic variations linked to drug resistance, they struggle with multiple resistance phenotypes and may yield false associations due to cross-resistance, necessitating validation through functional assays. Other statistical approaches, like those utilized by the World Health Organization (WHO), can mitigate artificial cross-resistance but require feature masking, prior knowledge, and expert rules. Thus, we introduce a novel Group Association Model (GAM) method to address these challenges in identifying genetic variations associated with drug resistance.
Materials and
Methods: To test GAM's efficacy, we utilized 7,179 Mtb isolates from the CRyPTIC database. These isolates underwent grouping, resulting in the creation of a control group and multiple drug-resistant groups with distinct profiles. Drug-resistant groups with < 2 isolates were excluded due to low statistical power. Non-synonymous genetic variants differing from the Mtb H37Rv reference genome were identified and analyzed for their association with drug resistance using Fisher’s exact tests, corrected for multiple comparisons. Variants not significantly enriched in any drug-resistant group were eliminated. Fisher’s exact test results were adjusted using Bonferroni corrections. Variants associated with resistance were identified based on odds ratios ≥1 and p-values (α*=0.05). Variants associated with drug-gene interactions beyond those detected by GAM analyses were identified using WHO confidence gating criteria.
Results, Conclusions, and Discussions: GAM successfully identified genetic variants associated with drug resistance in Mtb, demonstrating comparable performance across all analyzed drugs (Fig. 1a). GWAS gene-drug associations, reported using a linear mixed model (LMM), accurately identified gene targets but also produced multiple false-positive associations, particularly notable for rpoB (Fig. 1b). Additionally, spurious associations were detected with genes (pncA and rpsL) involved in resistance to drugs not analyzed in this dataset (pyrazinamide, and streptomycin), and with ethA, a reported target for ethionamide. GAM results excluded most cross-resistances detected by LMM, except for a single erroneous katG association with rifampin resistance, likely due to limited rifampin-sensitive but isoniazid-resistant groups. In comparison to WHO, GAM variants exhibited similar positive predictive values (PPV), demonstrating superior PPVs for amikacin and kanamycin (Fig. 1c). Specificity values showed no significant differences except in the case of rifampin (Fig. 1d). Sensitivity estimates were largely comparable, though GAM variants displayed slightly lower performance with isoniazid (Fig. 1e). Overall, both GAM and WHO methods achieved comparable results, but GAM's advantage lies in its independence from feature masking, prior knowledge, and expert rules, distinguishing it as a more streamlined and versatile approach. GAM accurately pinpoints genetic variants linked to resistance and minimizes false-positive associations, all without requiring prior knowledge. Thus, GAM could address the limitations of current drug resistance prediction methods to improve treatment decisions for drug-resistant microbial infections.
Acknowledgements (Optional): The work was primarily supported by research funding provided by the National Institutes of Health (W81XWH1910026, U01CA252965, R01AI144168, R01AI175618, R01HD090927, R01HD103511, R21NS130542).
