Predicting Azithromycin resistance in N. gonorrhoeae using Machine Learning.
Azithromycin resistance has a complex underlying genotype, involving numerous determinants. However, interest remains in predicting azithromycin susceptibility from genome sequence data, for example in culture-free diagnostic approaches. Machine learning offers a methodology to identify which genetic variants are most predictive of azithromycin resistance or susceptibility in N. gonorrhoeae, for automation within a digital tool.
Methods
In this study, we performed an extensive literature search to identify genetic determinants of azithromycin resistance for inclusion in the model. Subsequently, seven loci and 36 genetic variants were investigated across >15, 000 publicly available gonococcal genomes with azithromycin MIC data, via the PubMLST bacterial genetics database. A decision tree classifier was trained using 70% of these genomes to determine which genetic variants were most predictive of resistance/susceptibility, with the remaining 30% used in validation.
Results
Six predictive genotypic variations were included in the final model. These included three amino acid substitutions within the mtrCDE efflux pump operon, one nucleotide polymorphism in the mtrR promoter region, and two in the 23s rRNA gene. On the basis of these loci, the model was able to reliably predict azithromycin resistance, with an accuracy level of 95.3%. The positive predictive value for resistance was 78%, and for susceptibility 97%. A small number of susceptible isolates were falsely classified as resistant by the model, with 90% of these possessing MICs ranging from 0.5-1 µg/mL. This prediction tool has been made freely available on PubMLST, where all uploaded gonococcal genomes in which the key loci can be annotated will be assigned a deduced azithromycin resistance phenotype.
Summary
In conclusion, our machine learning approach can accurately predict azithromycin resistance based on only a small number of genetic variants. The implementation of this tool demonstrates how molecular approaches can be applied to facilitate culture-free detection of antimicrobial resistance in N. gonorrhoeae, even in the absence of a whole genome sequence. Future directions will include testing the robustness of the model in the face of new data, by examining its accuracy across diverse novel isolate datasets.