Testing every novel antibiotic combination for tuberculosis treatment shortening in the relapsing mouse model is impractical because the model is resource- and time-intensive. The authors aimed to build a computational model that predicts long-term relapse prevention from short-term measures obtained after 28 days (4 weeks) of treatment. Such a model would expand the number of regimens that can be screened and help prioritize regimens for further development.
Two short-term biomarkers were central to the work. The conventional measure of bacterial burden, lung colony forming units (CFU), was combined with an innovative biomarker, the ribosomal RNA synthesis (RS ratio), which characterizes drug effects on Mycobacterium tuberculosis health and activity. The RS ratio is intended to provide complementary information about bacterial physiological state beyond static viable counts.
Model development leveraged nine datasets comprising 58 unique treatment regimens. Across those datasets there were 843 short-term biomarker observations and 2,239 long-term relapse outcome observations. The datasets were used across three model-development iterations with external validations at each phase.
The modeling proceeded in three iterative stages. At each iteration the model was developed separately and then externally validated. Despite separate model development across iterations, the ultimate model structure did not require change to achieve optimal performance, indicating stability of the predictive relationships discovered.
Final model predictors included therapeutic indicators such as change from baseline in CFU and change in RS ratio. The model also incorporated corrections for experimental conditions to enable unbiased ranking of regimens across different experiments. These corrections were intended to adjust for variability introduced by trial conduct and allow comparisons of regimen performance between independent experiments.
External validation of the final model yielded an area under the receiver operator curve of 0.90, reflecting strong discriminative ability for predicting long-term relapse outcomes from 4-week biomarker data. The model successfully distinguished regimens with differing sterilizing durations (reported as 2-, 3-, and 4-month regimens) using only 28-day data.
The investigators evaluated model performance when removing either biomarker. They found that a CFU-only model performed similarly to a combined CFU and RS ratio model provided that the model accounted for the sterilizing contribution of individual drugs within regimens. The authors note that for new drugs where per-drug sterilizing contributions are not yet quantified, measurement of the RS ratio may be advantageous to predict relapse risk.
By predicting relapse from short-term measurements, the model can increase throughput for evaluating novel regimens in preclinical screening. It enables prioritization of regimens for resource-intensive long-term relapsing mouse experiments and supports selection of candidates for further development aimed at treatment shortening.
Because this report is a preprint, it has not undergone peer review. The manuscript discloses that two authors are co-inventors on a US patent related to the RS ratio; all other authors declared no competing interests. Funding sources reported in the article include the National Institutes of Health, the Gates Foundation, Gates Medical Research Institute, a CDC subcontract, and Veterans Affairs support. Specific methodological details, model parameters, and numerical effect estimates beyond those summarized here were not reported in the abstract.
A computational model combining 28-day CFU and RS ratio measurements, with correction for experimental factors and accounting for drug sterilizing contributions, can predict long-term relapse prevention in the relapsing mouse model. With an external-validation AUC of 0.90, the model differentiated between 2-, 3-, and 4-month regimens using only 4-week data. The approach can accelerate screening and prioritization of treatment-shortening regimens, and measurement of the RS ratio may be particularly useful when evaluating new drugs whose sterilizing contributions are unknown.