Quantitative modeling predicts mechanistic links between pre-treatment microbiome composition and metronidazole efficacy in bacterial vaginosis

Bacterial vaginosis is a condition associated with adverse reproductive outcomes and characterized by a shift from a Lactobacillus-dominant vaginal microbiota to a polymicrobial microbiota, consistently colonized by strains of Gardnerella vaginalis. Metronidazole is the first-line treatment; however, treatment failure and recurrence rates remain high. To understand complex interactions between Gardnerella vaginalis and Lactobacillus involved in efficacy, here we develop an ordinary differential equation model that predicts bacterial growth as a function of metronidazole uptake, sensitivity, and metabolism. The model shows that a critical factor in efficacy is Lactobacillus sequestration of metronidazole, and efficacy decreases when the relative abundance of Lactobacillus is higher pre-treatment. We validate results in Gardnerella and Lactobacillus co-cultures, and in two clinical cohorts, finding women with recurrence have significantly higher pre-treatment levels of Lactobacillus relative to bacterial vaginosis–associated bacteria. Overall results provide mechanistic insight into how personalized differences in microbial communities influence vaginal antibiotic efficacy.


nature research | reporting summary
April 2020 Field-specific reporting Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.

Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design
All studies must disclose on these points even when the disclosure is negative. The sample sizes for the experiments were not pre-determined statistically. The sample size of the in vitro monoculture data used for model parameterization which ranged from n = 3 for the parameterization of MNZ uptake and metabolism, n = 9 for the growth rates and carrying capacity, and n = 3-5 for sensitivity of the bacteria to MNZ. Completing these experiments in triplicate is typical in the field (Atassi, F., et al.  (2019)). As accuracy in parameterization was not the central goal of this manuscript the use of 3-9 independent biological replicates was satisfactory, and we additionally completed simulations over ranges of possible parameter values observed in the literature (Fig. 4, Supplementary Tables 2-3). Moreover, since we later validated the model that were based on measurements from monoculture data in co-culture, we further supported that the sample sizes and estimations from the monoculture were sufficient. We selected a larger sample size for the in vitro validation study ( (2017)). The de-identified clinical data sample size was limited to data previously collected in those studies.
Exclusion criteria were not pre-established for the in vitro data, but no data was excluded from in vitro experimental measurements. Data was excluded from the clinical cohorts based on the following predetermined (in regards to our statistical analysis) criteria (stated in the Methods text): 1) MNZ regimen was not completed; 2) Individual did not have BV according to Nugent scoring at the time of MNZ treatment; 3) Individual did not have follow-up data available; 4) Individual did not exhibit treatment failure as defined in the manuscript (resolved BV at an intermediate time point followed by a positive test for BV).
To verify the trends observed in the clinical data, we looked at two independent studies in distinct study populations. The in vitro data had a sample size of n = 18 co-cultures for each test condition. Additionally, when we observed variability in L. iners growth, we completed simulations to determine how growth dynamic variability influenced the model findings. All attempts at replication of model findings were successful.
The in vitro bacterial mono and co-cultures were not randomized, and no covariates are anticipated to influence the results as the experiments were all completed by the same individual and same setting. Clinical data was randomized as previously described in their respective publications ( Computational prediction of model findings by CYL was initially blinded to the experimental validation results by RKC, until experimenter confirmed the ratio dependent trends were observed in the data. Clinical data was blinded as previously described in their respective publications (Ravel et al., 2013 (PMID: 24451163) for the UMB-HMP data and Thurman et al., 2015 (PMID: 26204200) for the BV Conrad data).