Regression for accurate and sensitive grading of mutations diagnostic of antibiotic resistance in Mycobacterium tuberculosis

Sanjana G. Kulkarni, Sacha Laurent, Paolo Miotto, A Sarah Walker, Leonid Chindelevitch, Carl‐Michael Nathanson, Nazir Ismail, Timothy C. Rodwell, Maha FarhatmedRxiv 2024 · DOI

TL;DR

Researchers have developed a multivariate regression-based pipeline to identify antibiotic resistance mutations in Mycobacterium tuberculosis (TB). By analyzing over 50,000 bacterial isolates, this approach accounts for co-occurring mutations, identifying 29% more resistance-associated variants and increasing diagnostic sensitivity by an average of 3.2 percentage points compared to the current World Health Organization (WHO) standard.

What Problem Does This Paper Address?

Tuberculosis remains a leading cause of death from infectious agents globally. Rapidly identifying which drugs will work against a specific strain is critical for patient survival and preventing the spread of multi-drug resistant TB. Currently, the "SOLO" method used by the WHO relies on univariate associations—essentially looking at mutations in isolation.

The problem is that bacteria rarely have just one mutation. The SOLO method often excludes isolates with multiple candidate mutations to avoid "confounding," which means a massive amount of genomic data is thrown away. Furthermore, it struggles to distinguish between the primary mutation causing resistance and "compensatory" mutations that merely help the bacteria survive the fitness cost of being resistant.

Methods

The authors moved from univariate analysis to multivariate penalized (L2) logistic regression.

  • Multivariate Approach: Instead of looking at one mutation at a time, the model considers all non-silent mutations in a gene simultaneously. This allows the model to "tease apart" which specific mutation is driving the resistance phenotype when multiple variants are present.
  • L2 Regularization (Ridge Regression): Because TB evolves clonally, many mutations are "linked" (inherited together). L2 regularization helps handle this multicollinearity, preventing the model from becoming over-reliant on any single linked variant.
  • Population Structure Correction: To ensure associations weren't just artifacts of bacterial lineage (e.g., a mutation common in a specific geographic strain that happens to be resistant), the team used Principal Component Analysis (PCA) on over 6,000 neutral variants as covariates in their models.
  • Validation: They validated their binary (resistant/susceptible) results using linear regression on Minimum Inhibitory Concentration (MIC) data, which provides a more granular look at how much drug is needed to kill the bacteria.

Datasets

The study utilized a massive global collection of M. tuberculosis complex (MTBC) genomes:

  • WHO Dataset (N = 41,141): High-confidence phenotypes tested using WHO-approved methods.
  • ALL Dataset (N = 52,567): The WHO dataset plus additional results from other phenotypic approaches (like UKMYC 96-well plates).
  • The models covered 15 different antitubercular drugs, ranging from first-line staples like Isoniazid to newer "last-resort" drugs like Bedaquiline.

Key Insights and Ideas

  • Beyond "One Mutation, One Resistance": The model successfully identified compensatory mutations (specifically in the ahpC promoter). These don't cause resistance themselves but increase the expression of antioxidant enzymes to offset the damage caused by other resistance-conferring mutations in katG.
  • Identifying Hypersusceptibility: Surprisingly, the model found variants (Loss-of-Function in eis and mmpL5) that make bacteria more sensitive to certain drugs. This is a "hidden" insight that univariate methods usually miss but is vital for precision medicine and dosing.
  • Data Efficiency: By including isolates with multiple mutations, the regression approach "graded" (assigned a confidence level to) 218 more variants than the SOLO method. It essentially turned "noise" into "signal."
  • Automation Potential: The regression approach is highly scalable. It reduces the reliance on manual literature reviews and expert panels to "grade" mutations, paving the way for a more automated, data-driven update to global resistance catalogs.

Key Results

  • High Agreement: Regression detected 99% (452/457) of the resistance-associated variants already identified by the SOLO method, proving its reliability.
  • Increased Sensitivity: The regression-based catalog improved diagnostic sensitivity by an average of 3.2 percentage points (pp) across 15 drugs. For some drugs like Ethambutol, the gain was as high as 10.2 pp.
  • Minimal Trade-offs: This gain in sensitivity came with only a minor decrease in specificity (-1.0 pp) and positive predictive value (-1.8 pp).
  • Improved Balance: The F1 score (the harmonic mean of sensitivity and precision) averaged across all drugs improved from 69.2% (SOLO) to 70.1% (Regression).

Limitations and Future Work

  • The "Missing Data" Hurdle: Regression models cannot be trained on missing data. About 11% of variants couldn't be fit because they occurred in regions with poor sequencing coverage or were too rare (found in <5 isolates).
  • Correlation vs. Causality: While the model finds strong associations, it doesn't strictly prove causality. Some variants might still be "passengers" linked to true resistance markers, particularly in drugs with biased sampling like Bedaquiline.
  • Future Directions: The authors suggest that future models could incorporate better imputation for missing data and use more diverse global datasets to further refine the correction for population structure. There is also potential to use these models not just for grading mutations, but for direct clinical prediction of resistance from a raw genetic profile.

Summary updated

Abstract

Rapid genotype-based drug susceptibility testing for the Mycobacterium tuberculosis complex (MTBC) relies on a comprehensive knowledgebase of the genetic determinants of resistance. We built a catalog of resistance-associated mutations in MTBC using a novel regression-based approach and benchmarked it against the 2 nd edition of the World Health Organization mutation catalog. We trained multivariate logistic regression models on over 50,000 MTBC isolates to associate binary resistance phenotypes for 15 antitubercular drugs with variants extracted from candidate resistance genes. Regression detects 452/457 (99%) resistance-associated variants identified using the existing method ( a.k.a, SOLO method) and grades 218 (29%) more total variants than SOLO. The regression-based catalog achieves higher sensitivity on average (+3.2 percentage points, pp) than SOLO with smaller average decreases in specificity (−1.0 pp) and positive predictive value (−1.8 pp). The regression pipeline also detects isoniazid resistance compensatory mutations in ahpC and variants linked to bedaquiline and aminoglycoside hypersusceptibility. These results inform the continued development of targeted next generation sequencing, whole genome sequencing, and other commercial molecular assays for diagnosing resistance in MTBC. In addition to grading genetic variants by their associations with phenotype, regression models could potentially provide an accurate and scalable method of predicting antibiotic resistance from bacterial genetic profiles.
BibTeX
@misc{kulkarni2024regression,
  author = {Sanjana G. Kulkarni and Sacha Laurent and Paolo Miotto and A Sarah Walker and Leonid Chindelevitch and Carl‐Michael Nathanson and Nazir Ismail and Timothy C. Rodwell and Maha Farhat and Maha R. Farhat},
  title = {{Regression for accurate and sensitive grading of mutations diagnostic of antibiotic resistance in <i>Mycobacterium tuberculosis</i>}},
  year = {2024},
  doi = {10.1101/2024.07.01.24309598},
  howpublished = {medRxiv},
}
Figure