Selection and prioritization of candidate combination regimens for the treatment of tuberculosis

Accelerated tuberculosis drug discovery has increased the number of plausible multidrug regimens. Testing every drug combination in vivo is impractical, and varied experimental conditions make it challenging to compare results between experiments. Using published treatment efficacy data from a mouse tuberculosis model treated with candidate combination regimens, we trained and externally validated integrative mathematical models to predict relapse in mice and to rank both previously experimentally studied and unstudied regimens by their sterilization potential. We generated 18 datasets of 18 candidate regimens (comprising 11 drugs of six classes, including fluoroquinolone, nitroimidazole, diarylquinolines, and oxazolidinones), with 2965 relapse and 1544 colony-forming unit (CFU) observations for analysis. Statistical and machine learning techniques were applied to predict the probability of relapse in mice. The locked down mathematical model had an area under the receiver operating characteristic curve (AUROC) of 0.910 and showed that bacterial kill measured by longitudinal CFU cannot account for relapse alone and that sterilization is drug dependent. The diarylquinolines had the highest predicted sterilizing activity in the mouse model, and the addition of pyrazinamide to drug regimens provided the shortest estimated tuberculosis treatment duration to cure in mice. The mathematical model predicted the effect of treatment combinations, and these predictions were validated by conducting 11 experiments on previously unstudied regimens, achieving an AUROC of 0.829. We surmise that the next generation of tuberculosis drugs are highly effective at treatment shortening and suggest that there are several promising three- and four-drug regimens that should be advanced to clinical trials.