Genomic tool SensiTyper predicts gonorrhoea drug susceptibility and tailors treatment from whole-genome sequences

A rule-based computational tool called SensiTyper can predict which antibiotics will work against individual gonorrhoea infections directly from whole-genome sequencing data, and translate those predictions into susceptibility-guided treatment recommendations, according to an analytical validation study published in Nature Communications. The tool was applied to 13,701 Neisseria gonorrhoeae genomes sampled between 1928 and 2024 from 68 countries, including cultured isolates and clinical molecular samples, and its predictions were benchmarked against conventional phenotypic antimicrobial susceptibility testing.

The motivation is pressing. Neisseria gonorrhoeae, the bacterium that causes gonorrhoea, has developed resistance to every antimicrobial introduced for treatment since the mid-1930s, including the last remaining first-line option, the extended-spectrum cephalosporin ceftriaxone. This has produced a stable population of multidrug-resistant strains in many countries, sporadic ceftriaxone treatment failures worldwide, and, in the most recent decade, extensively drug-resistant (XDR) strains resistant to both ceftriaxone and azithromycin, the backbone of current first-line therapy. International guidelines therefore still recommend empiric therapy, typically high-dose ceftriaxone alone or ceftriaxone plus azithromycin, chosen without knowing a particular infection’s susceptibility profile.

SensiTyper departs from earlier genomic prediction efforts by applying transparent, biologically interpretable rules rather than statistical inference or machine learning. Paired-end sequencing reads are analysed with the software ARIBA against curated reference databases of antimicrobial resistance determinants, and antimicrobial-specific decision rules then classify each isolate as susceptible or non-susceptible to six drugs: ceftriaxone, azithromycin, ciprofloxacin, spectinomycin, zoliflodacin and gepotidacin. The rules account for single-nucleotide substitutions, gene disruptions and mosaic alleles, drawing on established resistance mechanisms such as mosaic penA alleles encoding penicillin-binding protein 2 alterations for ceftriaxone, 23S rRNA mutations and mosaic MtrD efflux variants for azithromycin, and gyrA and parC mutations, especially GyrA S91F, for ciprofloxacin.

A second module, SensiTreat, converts the susceptibility profile into a stepwise treatment recommendation. Antimicrobial regimens are recommended only for isolates predicted to be susceptible, and ineffective options are excluded sequentially until a treatable option is found. Users define the priority order of available antimicrobials, which allows different strategies: a guideline-aligned order starting with ceftriaxone monotherapy, a dual-therapy strategy beginning with ceftriaxone plus azithromycin, or an antimicrobial stewardship configuration that prioritises ciprofloxacin when genomic susceptibility is predicted. Isolates that cannot be assigned any regimen are automatically flagged as XDR alerts for expert follow-up, test-of-cure and comprehensive contact tracing.

The predictive performance was strong for most drugs. Susceptibility predictions for ceftriaxone, ciprofloxacin, spectinomycin and zoliflodacin achieved sensitivities and specificities of 98.9 to 100 percent. For ceftriaxone, sensitivity was 100 percent and specificity 98.9 percent, with a Matthews correlation coefficient of 0.994. Ciprofloxacin predictions reached 99.5 percent sensitivity and 99.8 percent specificity. Spectinomycin and zoliflodacin predictions showed complete concordance with phenotypic testing, although nearly all isolates were susceptible to these two drugs, so the exceedingly low number of resistant isolates produced wide confidence intervals for specificity, negative predictive value and very major error rates.

Azithromycin proved harder. Susceptibility predictions had 92.5 percent sensitivity and 97.1 percent specificity, but a markedly lower negative predictive value of 52.2 percent, indicating that azithromycin-susceptible isolates were often misclassified as resistant. Thirty-one azithromycin-resistant isolates were predicted susceptible; their minimum inhibitory concentrations were 2 mg/L in 27 cases, 4 mg/L in three and 8 mg/L in one, with wild-type MtrD and 23S rRNA, and the isolates were not available for retesting. The authors attribute the reduced performance to the polygenic and sometimes subtle mechanisms of decreased azithromycin susceptibility, with discrepancies concentrated among isolates with MICs near the epidemiological cut-off value that are difficult to classify even by phenotypic testing.

Gepotidacin exposed a different limitation. Sensitivity was 100 percent and the positive predictive value 99.8 percent, but specificity was 0 percent and the very major error rate 100 percent: six isolates resistant to gepotidacin under the recently stated US FDA resistance breakpoint of MIC at or above 4 mg/L were all predicted susceptible. The reason is that these isolates with low-level phenotypic resistance do not carry the only known resistance determinants in N. gonorrhoeae, the concomitant GyrA A92T and ParC D86N substitutions. The authors note that the FDA breakpoints are temporary or provisional interpretive criteria, not yet official clinical breakpoints from EUCAST or the US CLSI, and that in the microbiologically evaluable population of the gepotidacin phase 3 randomized trial no failures to treat urogenital (0 of 187) or anorectal (0 of 26) gonorrhoea were detected, although two of 16 pharyngeal cases were not cured.

In the treatment-recommendation evaluation, a susceptibility-guided strategy prioritising ceftriaxone monotherapy assigned ceftriaxone 1 g to 13,286 of 13,633 isolates, a coverage fraction of 97.5 percent, with 100 percent sensitivity and a very major error rate of 1.1 percent, corresponding to four misclassifications. Azithromycin 2 g monotherapy was assigned to 257 of the remaining 347 evaluable isolates, spectinomycin to 13 isolates with full concordance, and 77 isolates received no recommendation because spectinomycin MIC data were lacking; these were designated XDR alert isolates. When zoliflodacin and gepotidacin were included as additional options, all 77 XDR isolates were susceptible based on available MICs and correctly predicted susceptible, so no isolate remained without a treatment option. Under the stewardship configuration, 4,144 of 11,741 isolates, 35.3 percent, were directly assigned ciprofloxacin with 99.5 percent sensitivity, and subsequent ceftriaxone-based steps maintained high accuracy. SensiTreat also recommended ceftriaxone 1 g monotherapy for all 68 NAAT-derived genomes with 100 percent concordance with matched cultured isolates.

The study has limitations the authors state explicitly. Sex, gender and age of patients were unavailable; although isolates came from 68 countries, a large proportion were from the WHO European Region, causing suboptimal global representativeness, and reliance on MIC-tested cultured isolates further skewed the dataset. Clinical outcome data were not available, so a prospective clinical validation linking prediction, treatment recommendation and outcome would be valuable, and regulatory approvals may be required in some settings. Cefixime, penicillin and tetracycline were excluded because they are no longer recommended in most countries. The correlations between in vitro resistance and clinical cure are also suboptimal for several drugs; for example, infections caused by most ceftriaxone-resistant strains with MICs of 0.25 to 0.5 mg/L, especially urogenital and anorectal infections, can be effectively cured with high-dose 1 g ceftriaxone therapy.

SensiTyper is already implemented in the authors’ clinical routine diagnostic workflow in Sweden, where complete resistance prediction and treatment recommendations are obtained in roughly 1.5 to 2 minutes for ten genomes with an average depth of 50 to 100 times on a conventional laptop. The tool is freely available and open source. For its full clinical utility, however, whole-genome sequencing directly from clinical NAAT samples, which most patients receive as their diagnosis, will need to be implemented in routine laboratories; the authors suggest that rapid real-time long-read sequencing approaches are likely to make this feasible soon. They also caution that treatment recommendations must consider anatomical site, since spectinomycin, zoliflodacin and gepotidacin show suboptimal efficacy against pharyngeal gonorrhoea, and that the tool can equally serve public health surveillance by providing AMR prediction data for local, national and international monitoring and evidence-based treatment policy.

Subject of Research: Analytical validation of a rule-based genomic tool that predicts antimicrobial susceptibility and individualized treatment for gonorrhoea

Article Title: Genomic surveillance meets clinical practice: rule-based individualized treatment prediction for gonorrhoea using SensiTyper

Article References: Golparian, D., Sánchez-Busó, L., & Unemo, M. (2026). Genomic surveillance meets clinical practice: rule-based individualized treatment prediction for gonorrhoea using SensiTyper. Nature Communications, 17(1), Article 11293. https://doi.org/10.1038/s41467-026-79037-z

Image Credits: AI Generated

DOI: 10.1038/s41467-026-79037-z

Keywords: Neisseria gonorrhoeae, SensiTyper, whole-genome sequencing, antimicrobial resistance, ceftriaxone, azithromycin, ciprofloxacin, zoliflodacin, gepotidacin, treatment recommendations, genomic surveillance, antimicrobial stewardship